Abstract
Background
Medical data interpretation traditionally relies on healthcare professionals as intermediaries, which can limit patient autonomy and engagement. Large Language Models (LLMs) present an opportunity to transform this paradigm by enabling direct patient access to AI-generated interpretations; however, comparative research on their effectiveness across different medical data types and communication modalities remains limited. This study explores how direct LLM-augmented interpretation of medical data, in which patients use an AI system to receive real-time explanations of laboratory and radiological results, compares with healthcare professional-led interpretation across different data modalities, with particular attention to patient comprehension, empowerment, and technology acceptance.
Methods
Using a mixed-methods approach with a within-subjects experimental design, 45 demographically diverse participants experienced six scenarios: blood work and medical imaging interpretations delivered via (1) healthcare professional phone consultation, (2) in-person consultation, or (3) LLM interaction through a custom-configured ChatGPT-4o interface (Medical Explainer AI) designed to provide plain-language explanations of findings, highlight abnormal values, contextualize clinical significance, explain medical terminology, and adapt explanation complexity based on user feedback.
Results
LLM interaction significantly enhanced diagnostic comprehension (mean difference = 1.3 compared to phone consultation, p < 0.001), reduced cognitive load, increased perceived control, and improved time efficiency. Healthcare professional-led interpretation, particularly in-person, maintained advantages in fostering trust, reducing anxiety, and enhancing confidence in decision-making. The benefits of LLM interaction were more pronounced for blood work than for medical imaging interpretation. Age, education level, and health literacy significantly moderated the effectiveness of different interpretation methods.
Conclusions
LLMs offer complementary rather than replacement capabilities for medical data interpretation, excelling in enhancing comprehension, control, and efficiency, while healthcare professionals provide superior relational value through trust, confidence, and emotional support. Implementation strategies should leverage the strengths of both approaches, carefully considering data complexity and patient characteristics to maximize benefits while ensuring equitable access.
Supplementary Information
The online version contains supplementary material available at https://doi.org/10.1186/s12911-026-03686-9.
Keywords: Large Language Models (LLMs), Medical data interpretation, Human-AI interaction, Patient empowerment, Technology acceptance, Healthcare communication
Background
Medical data interpretation, including laboratory results and diagnostic imaging, represents a critical interface in healthcare delivery where clinical expertise meets patient understanding. Traditionally, healthcare professionals have served as the exclusive intermediaries in this process, translating complex medical information into actionable insights for patients through phone consultations or in-person meetings [1]. However, this approach often creates information asymmetry, potentially diminishing patient autonomy and engagement in their own healthcare decision-making [2, 3]. The rapid advancement of artificial intelligence (AI), particularly Large Language Models (LLMs) such as ChatGPT, Claude, and Google Gemini, presents an unprecedented opportunity to transform this paradigm [4]. These models demonstrate sophisticated natural language understanding and generation capabilities, potentially enabling patients to independently interpret their medical data or augment professional consultations [5]. As healthcare systems globally contend with resource constraints, workforce shortages, and increasing service demands, LLMs may offer a scalable, accessible solution to enhance patient access to timely and comprehensible medical information interpretation [6]
Despite their potential, integrating LLMs into medical data interpretation raises significant questions about efficacy, accuracy, patient experience, and ultimate health outcomes [7, 8]. A critical review of existing literature reveals a substantial research gap: while preliminary investigations have explored LLM applications in clinical decision support systems [9, 10], empirical research examining their role in patient-facing interpretation of actual medical data remains notably limited. Comparative studies are particularly lacking in evaluating LLM-mediated interpretation against traditional healthcare professional-led approaches across different data modalities and communication channels. This gap is especially significant given the rapid deployment of these technologies in healthcare settings without corresponding evidence regarding their comparative effectiveness, patient acceptability, or psychological impact.
The theoretical underpinnings for understanding this phenomenon lie at the intersection of three complementary frameworks: technology acceptance, patient empowerment, and information processing. The Unified Theory of Acceptance and Use of Technology (UTAUT) provides a robust conceptual foundation for understanding the determinants of patient adoption of LLM-based interpretation technologies, including performance expectancy, effort expectancy, social influence, and facilitating conditions [11]. Moreover, patient empowerment theory offers a lens through which to examine how increased access to interpretable medical information may enhance patients’ perceived control, decisional involvement, and self-efficacy in health management [12]. Finally, information processing theory complements these perspectives by elucidating how different presentation modalities and interaction patterns might affect cognitive load, comprehension, and retention of complex medical information [13]. This study suggests that integrating these theoretical frameworks enables a multidimensional analysis of how different interpretation modalities influence patient experiences and outcomes.
This study addresses critical empirical and theoretical gaps by systematically comparing patient experiences across six distinct scenarios of medical data interpretation, representing the factorial combination of two medical data types (blood work and medical imaging) and three interpretation methods (healthcare professional via phone, healthcare professional in-person, and LLM interaction). Through examination of key variables derived from our integrated theoretical framework, including diagnostic comprehension, perceived control, information satisfaction, decision confidence, anxiety reduction, and behavioral intention, we attempt to address the fundamental research question: How does the integration of LLMs in medical data interpretation affect patient experiences, comprehension, psychological outcomes, and adoption intentions compared to traditional healthcare professional-led approaches?
The selection of blood work and medical imaging as the two focal data modalities is strategically grounded in their contrasting characteristics. Blood work represents structured, quantitative medical data, discrete numerical values with clear reference ranges, and standardized interpretation frameworks. This structured nature creates favorable conditions for LLM interpretation through pattern matching and explanation. In contrast, medical imaging provides visual, qualitative data that require spatial reasoning, pattern recognition, and interpretive judgment to identify radiological findings. This modality presents substantially greater interpretive complexity for LLMs, requiring visual understanding, integration of multiple perspectives, and navigation of diagnostic uncertainty. By examining both modalities, we can systematically investigate whether LLMs’ comparative advantages vary with data structure and complexity. We hypothesize that LLMs demonstrate superior performance for structured numerical data but have more limited capabilities for complex visual interpretation. This contrast strengthens our ability to identify boundary conditions under which LLM-augmented interpretation is most beneficial and when professional human expertise remains essential.
The significance of this investigation extends beyond academic inquiry, addressing pressing challenges in healthcare delivery. Theoretical contributions include: 1. development of an integrated conceptual framework that synthesizes technology acceptance, patient empowerment, and information processing perspectives in the context of AI-augmented healthcare communication; 2. empirical validation of pathways through which interpretation modalities influence cognitive, psychological, and behavioral outcomes; and 3. identification of boundary conditions and moderating factors that influence the effectiveness of different interpretation approaches. Practical contributions encompass:1. Evidence-based guidance for healthcare systems on optimal integration of LLM technologies in patient-facing medical data interpretation services; 2. identification of specific contexts and data types where LLM augmentation offers maximal benefit; 3. delineation of patient segments most likely to benefit from different interpretation modalities; and 4. concrete recommendations for LLM interface design and implementation that maximize comprehension, empowerment, and acceptance while mitigating potential limitations.
As healthcare systems globally navigate the complex intersection of technological advancement, resource constraints, and patient-centered care imperatives, this research provides critical insights to inform the strategic integration of AI technologies to enhance healthcare accessibility, efficiency, and patient empowerment without compromising the essential human dimensions of medical communication. By empirically examining how different interpretation modalities shape patient experiences and outcomes, this study contributes to the development of more effective, equitable, and patient-centered approaches to medical data interpretation in an increasingly digital healthcare ecosystem.
Literature review and theoretical framework
Medical data interpretation and patient engagement
Historically, medical data interpretation (ranging from lab results to imaging and clinical notes) has been the exclusive domain of healthcare professionals [14]. Physicians, nurses, and allied health staff have acted as gatekeepers, translating technical jargon and complex findings into recommendations and treatment plans that patients can understand and act upon. This intermediary role was essential due to the complexity of medical information and the lack of accessible tools for patients to interpret such data themselves [15]. However, the rise of data analytics, digital engagement platforms, and patient decision support tools is reshaping this dynamic. The clinician’s role is evolving from sole interpreter to collaborative partner, supported by data analysts and empowered patients [16]. As healthcare continues to move toward personalized, data-driven, and patient-centered models, effective medical data interpretation and meaningful patient engagement will be increasingly shared responsibilities among clinicians, technical experts, and patients.
Moreover, the healthcare landscape has undergone a significant transformation in recent years, with growing recognition of the importance of patient engagement and empowerment in medical decision-making [17]. Healthcare professionals have become increasingly aware of the important role of patient involvement and patient preference in healthcare. Thus, patient engagement has evolved from a model in which patients were passive recipients of information to one emphasizing active partnership [18]. Patient engagement involves informing patients about their health and treatment, empowering them to participate in decision-making, and improving outcomes [19]. Engaged patients are more likely to adhere to treatment plans and experience better health results. Technologies such as patient portals, mobile apps, and decision aids provide patients with tailored information, helping them make informed choices. These tools operationalize patient engagement by delivering relevant data in accessible formats, fostering shared decision-making. Overcoming language and cultural barriers through translation services and multilingual digital platforms further enhances patient engagement, especially in diverse or cross-border healthcare settings [20].
Integrating advanced analytics and digital engagement platforms has transformed patient engagement in many ways. The systematic data collection and analysis of patient interactions enable healthcare systems to understand behaviors, preferences, and outcomes, leading to more personalized care [21]. Digital platforms generate significantly more actionable data points per interaction than traditional methods, thereby enhancing the quality and relevance of insights provided to patients. Digital platforms now serve as central resources with advanced data-collection capabilities, enabling the aggregation of formerly highly fragmented data. Analytics incorporating data from multiple sources (mobile, web, in-person, telephone) outperform single-channel approaches, offering a more comprehensive view of patient needs and behaviors [22]. The processing and analysis capabilities of these digital platforms enable patients to monitor their health journey in real time, thereby actively engaging in care processes. For instance, mobile applications allow patients to access and share their health records directly, increasing their autonomy and facilitating more meaningful engagement with providers [23].
Despite these advances, challenges remain in fully implementing patient engagement approaches. For example, implementing robust patient engagement strategies across organizations can be difficult, requiring standardized tools and processes developed in partnership with patients [24]. Also, healthcare organizations must invest in training and capacity-building to integrate patient-centered thinking and data-driven approaches across their operations [25]. Healthcare providers are concerned about the quality (accuracy, reliability, and timeliness) of medical data available to patients via digital tools [26]. Finally, ensuring that digital tools are accessible to all patients, regardless of language, literacy, or technological proficiency, remains a critical concern [27]. Therefore, digital transformation in healthcare has accelerated the shift toward greater patient engagement; however, technological challenges persist in fostering dynamic collaboration between healthcare professionals and patients, and there remains significant room for improvement.
AI and LLMs in healthcare
The application of artificial intelligence in healthcare has expanded rapidly, particularly in diagnostic support systems and clinical decision-making tools [28]. LLMs represent a significant advancement in AI capabilities, demonstrating strong potential to transform healthcare by streamlining administrative and clinical workflows, enhancing patient engagement, and supporting more informed clinical decision-making [29]. Integrating AI, particularly LLMs, into healthcare transforms how medical data, such as lab results, imaging reports, and clinical notes, are analyzed and communicated to patients. Early studies of LLMs in healthcare contexts have shown promising results. He et al. (2024) demonstrated that large language models could generate responses to patient questions about laboratory results with quality comparable to peer users, though with important limitations in handling complex or ambiguous cases [5]. Similarly, a study found that LLMs could effectively explain radiological findings to simulated patients, though the explanations sometimes lacked the trace of human experts [30].
LLMs such as ChatGPT and Claude 3.5 Sonnet are increasingly being deployed to interpret and simplify medical information [31]. Stanford Health Care’s AI tool drafts plain-language explanations of lab and imaging results for patients, which clinicians review before sharing [32]. This reduces administrative burdens while ensuring clarity and accuracy. Recent studies show LLMs can summarize radiology reports and discharge documents into patient-friendly formats, improving accessibility [33]. For example, ChatGPT-4 enhanced the readability of musculoskeletal radiology reports, though occasional oversimplifications were noted [34]. A study suggests that LLMs with reasoning capabilities could be used in clinical settings such as diabetes management [35]. LLMs are reshaping patient-provider interactions through personalized education as models like GPT-4 answer patient questions with 90%+ accuracy on topics ranging from hypertension to post-operative care, though readability remains a barrier for low-literacy populations [34]. Also, healthcare AI agents send medication reminders, monitor symptoms via chatbots, and triage patients (e.g., Buoy Health) [36]. These tools democratize access to medical knowledge, empowering patients to participate in shared decision-making.
Despite these advances, performance variability persists. While some LLMs outperform human responses in accuracy and relevance for lab result questions [34], others struggle with contextual details, risking omissions or hallucinations [37]. Moreover, over-reliance on AI-generated content risks the spread of misinformation if not rigorously validated. Studies highlight discrepancies in LLM outputs, such as plausible but incorrect interpretations of lab results or incomplete discharge summaries [38]. While LLMs improve access to information, disparities persist for non-English speakers and those with limited digital literacy [39]. Training data biases and data security risks necessitate robust governance frameworks to protect patient confidentiality [40]. Clinicians emphasize the need for transparent AI decision-making and ongoing human oversight to mitigate these risks.
Thus, recent research indicates that LLMs can hold immense potential to enhance medical data interpretation and patient engagement by translating complex information into actionable insights. While challenges such as accuracy gaps and ethical concerns persist, the collaborative integration of AI with clinician expertise promises to advance personalized, equitable care. Despite existing literature, significant gaps remain in understanding how LLMs compare with healthcare professionals in real-world medical data interpretation scenarios, particularly from the patient perspective. The limited comparative research available has primarily focused on either general metrics, such as accuracy and readability [41] or physicians’ perspectives rather than patient experience variables [42].
Theoretical framework
This study integrates three complementary theoretical frameworks to comprehensively examine LLM-augmented interpretation of medical data. These theories contribute to a deeper understanding of how the general public may choose to engage with medical question answering using LLMs. This section highlights these three theories.
Unified theory of acceptance and use of technology (UTAUT)
The UTAUT, developed by Venkatesh, Morris, Davis, and Davis [11] provides a robust framework for understanding technology adoption across industries, including healthcare. The theory identifies key constructs influencing behavioral intention to use technology, including performance expectancy, effort expectancy, social influence, and facilitating conditions. UTAUT has been applied in healthcare settings to understand patient adoption of various technologies, from electronic health records to telemedicine platforms [43, 44]. UTAUT offers explanatory power for this study regarding how patients might accept and adopt LLM-based interpretation tools compared to traditional healthcare professional consultation. First, LLMs represent a novel technology with which most patients have limited familiarity in healthcare contexts, creating a classic technology adoption scenario where perceived benefits must outweigh perceived risks and effort [4, 45]. Second, the model’s emphasis on performance expectancy aligns precisely with the critical nature of medical data interpretation, where accuracy and comprehensiveness directly impact health decisions and outcomes [7]. Third, UTAUT’s construct of effort expectancy captures the potential advantage of LLMs in providing on-demand, conversational interpretation experiences compared to the scheduling and access barriers of traditional healthcare consultations [46].
Variables derived from this framework include performance expectancy, effort expectancy, and behavioral intention to use. These constructs enable empirical assessment of how different interpretation modalities influence both the cognitive evaluation of the technology’s value proposition and the ultimate behavioral intention to adopt LLM-based interpretation in future medical data interpretation scenarios. In operationalizing UTAUT for this study, we leveraged its well-established technology acceptance constructs to examine the adoption factors of LLMs as emerging healthcare technologies, recognizing that understanding user acceptance is fundamental to the successful implementation of any novel technological intervention in clinical settings
Patient empowerment theory
Patient empowerment represents a process through which individuals gain greater control over health decisions [47, 48]. This theoretical framework emphasizes the importance of knowledge, skills, and confidence in enabling patients to participate actively in their healthcare. Small, Bower [49] conceptualized patient empowerment as involving multiple dimensions, including perceived control, psychological coping, and self-efficacy. Patient empowerment theory provides a crucial perspective for our study, as it directly addresses the psychological and behavioral dimensions of how patients engage with their medical information, an aspect that technological frameworks such as UTAUT may not fully capture. This theory is particularly relevant to our research on LLM-mediated medical interpretation because it enables us to examine how different interpretive modalities may influence patients’ sense of agency and control over their health information [50]. The growing emphasis on patient-centered care in contemporary healthcare systems necessitates understanding not only whether patients will adopt a technology, but also how that technology affects their participation in healthcare decision-making processes [51].
LLMs offer a unique opportunity to enhance patient empowerment through democratizing access to comprehensible medical information. Unlike traditional healthcare communication models, where healthcare professionals control information flow, LLM-augmented interpretation potentially allows patients to explore information at their own pace, ask clarifying questions without time constraints, and revisit explanations as needed [52]. These capabilities directly address key barriers to patient empowerment identified in previous research, including limited consultation time, unaddressed information needs, and difficulty understanding medical terminology [53]. In the context of medical data interpretation, patient empowerment theory suggests that increased access to interpretable information may enhance patients’ sense of control and engagement. Unlike technological adoption theories focusing on whether patients will use a system, patient empowerment theory examines how information access transforms patients’ relationships with their healthcare journey and decisions [54]. This perspective is essential for evaluating the full impact of LLM-augmented interpretation beyond simple usage metrics. Variables derived from this framework include perceived control, decision-making confidence, and self-efficacy in understanding medical information. These constructs allow us to empirically assess the psychological impact of different interpretation modalities on patients’ sense of agency and capability, complementing the technology adoption variables from UTAUT to provide a more comprehensive understanding of patient experiences with LLM-augmented medical data interpretation.
Information processing theory
Information processing theory, originating from cognitive psychology, provides insights into how individuals acquire, process, and retain information [55]. This framework is particularly relevant to understanding how different presentation modalities might affect comprehension and retention of complex medical information. Mayer’s (2005) cognitive theory of multimedia learning, an extension of information processing theory, suggests that people learn more deeply from words and pictures than from words alone, and that the simultaneous presentation of explanatory words and pictures yields better learning than successive presentation [56]. This theoretical perspective is essential to our study because it addresses the cognitive dimensions of how patients interact with and understand medical data interpretations, a critical aspect that neither technology acceptance nor empowerment theories fully capture. Information processing theory provides a framework for examining how the structural and presentational aspects of information delivery across different modalities (e.g., healthcare professional via phone, in person, or LLM interaction) affect patients’ cognitive processing [57, 58].
Medical data interpretation inherently involves processing complex, often technical information that exceeds the working memory capacity of most individuals without medical training [59]. The cognitive load construct from information processing theory is particularly valuable for our research as it allows us to assess how different interpretation modalities might reduce extraneous cognitive load, the mental effort imposed by the way information is presented rather than the inherent complexity of the content itself [59]. LLM-based interpretation potentially offers advantages in this domain through features such as conversational interaction patterns, the ability to adjust explanation complexity based on user feedback, and the capacity to revisit concepts from multiple angles [60, 61]. The theory’s emphasis on information chunking and organization also aligns with how different interpretation methods structure and deliver medical information. LLMs can dynamically organize information into manageable segments based on user queries, potentially optimizing the cognitive processing requirements compared to the more linear information delivery often characteristic of time-constrained healthcare professional consultations [62]. This capability may be particularly relevant for patients with varying levels of health literacy or cognitive abilities, addressing a key dimension of healthcare equity.
Our incorporation of information processing theory allows us to examine not only whether patients accept or feel empowered by different interpretation modalities, but also how effectively they process and comprehend the information presented. This cognitive dimension is crucial for truly effective medical data interpretation, as even technologies that are readily adopted and promote feelings of empowerment ultimately fail if they do not support an accurate understanding of medical information [63]. Variables derived from this framework include information comprehension, information satisfaction, and cognitive load. When measured alongside the technology acceptance and patient empowerment variables, these constructs provide a comprehensive assessment of how different interpretation modalities affect the full spectrum of patient experiences, from cognitive processing to psychological responses to behavioral intentions.
To operationalize information processing theory in this study, we focus on three specific and complementary aspects: diagnostic comprehension, information satisfaction, and cognitive load. Diagnostic comprehension operationalizes the cognitive acquisition component of information processing, specifically, how effectively patients understand the substantive medical information presented through different modalities. This measure captures Sweller’s [55] conceptualization of how task-relevant information is encoded into working memory. Information satisfaction operationalizes the subjective evaluation dimension of information processing, reflecting whether the information delivered meets patients’ informational needs and expectations [56]. Unlike comprehension (which focuses on objective understanding), satisfaction captures the phenomenological experience of receiving information, addressing both content adequacy and presentation quality [57]. Finally, cognitive load operationalizes the cognitive resource-demand dimension, specifically the extraneous cognitive load imposed by the presentation modality rather than the content’s inherent complexity [59]. By measuring these three dimensions concurrently, we capture the multifaceted nature of information processing: what patients learn (comprehension), how satisfied they are with the learning experience (satisfaction), and the mental effort required (cognitive load). This three-dimensional operationalization enables us to examine whether interpretation methods that reduce cognitive burden (such as LLMs’ conversational approach) can simultaneously enhance comprehension and user satisfaction, or whether trade-offs exist between these dimensions.
Integrated conceptual model
The study’s integrated theoretical framework, illustrated in Fig. 1, synthesizes three complementary perspectives to explain the complex processes underlying experiences of medical data interpretation. Information processing theory provides the cognitive foundation, explaining how interpretation methods influence comprehension, satisfaction, and cognitive load during medical information processing. This cognitive processing directly informs patient empowerment theory, which posits that enhanced understanding of information contributes to greater perceived control, decision confidence, and health self-efficacy (psychological states crucial for active patient engagement). These empowerment constructs, in turn, influence technology acceptance variables derived from UTAUT, including performance expectancy, effort expectancy, and behavioral intention to use. Importantly, the framework incorporates the moderating influences of demographic variables (age, education, health literacy), while medical data type (blood work versus medical imaging) serves as a second within-subjects independent variable, accounting for the differential effectiveness of interpretation methods across participant characteristics and data complexity, thereby providing a comprehensive explanatory model for understanding LLM-augmented versus traditional healthcare professional-led interpretation experiences.
Fig. 1.
Integrated framework
By integrating these three theoretical frameworks, we propose a comprehensive model for understanding the comparative effects of different medical data interpretation scenarios. This model suggests that the method of interpretation (healthcare professional via phone, healthcare professional in-person, or LLM interaction) affects information processing variables (comprehension, satisfaction, cognitive load), which in turn influence patient empowerment variables (perceived control, decision confidence, self-efficacy), ultimately affecting technology acceptance variables (performance expectancy, effort expectancy, and behavioral intention to use). This integrated theoretical approach allows us to examine not only whether patients accept LLM-augmented interpretation, but also how different interpretation modalities affect the cognitive and psychological processes underlying the understanding of medical data and the making of healthcare decisions.
Methods
Research design
This study employs a within-subjects experimental design in which each participant experiences all six interpretation scenarios in a randomized order to control for order effects. The six scenarios represent combinations of two medical data types (blood work results and medical imaging) and three interpretation methods (a healthcare professional via phone, a healthcare professional in person, and an LLM interaction). Thus, this study uses a 3 × 2 factorial experimental design to examine the effects of communication modality and medical data type on patient perceptions and outcomes. Table 1 depicts the factorial experimental design used in this study.
Table 1.
3 × 2 factorial experimental design
| Independent Variable | Levels | Description |
|---|---|---|
| Communication Modality | Healthcare professionals via phone | Telephone consultation with a board-certified physician who explains and interprets the medical data |
| Healthcare professional in-person | Face-to-face consultation with a board-certified physician who explains and interprets the medical data | |
| LLM-mediated consultation | Interaction with a custom-configured GPT-4o model through a chat interface for medical data interpretation | |
| Medical Data Type | Blood work analysis | Interpretation of standardized blood work results, including complete blood count (CBC), comprehensive metabolic panel (CMP), lipid panel, and thyroid function tests |
| Medical imaging interpretation | Interpretation of standardized knee/elbow X-ray images with subtle but clinically relevant findings of varying complexity |
Therefore, our research design yields the following six scenarios, as shown in Fig. 2:
Fig. 2.
Six experimental conditions
Participants, sampling, recruitment, and ethics
The study employed a rigorous stratified random sampling methodology to recruit 45 participants, ensuring appropriate representation across key demographic variables, including age, gender, education level, and prior healthcare experience. This purposive stratification approach was essential for capturing heterogeneous perspectives on medical data interpretation modalities and for addressing potential demographic moderating effects identified in previous technology acceptance research [64]. Inclusion criteria were established as: 1. adults aged 20–70 years; 2. English language proficiency sufficient for complex health communication; 3. basic digital literacy enabling independent technology interaction; 4. experience with receiving at least one medical test result within the previous 24 months; and 5. current residence within Miami-Dade and Broward County (in Florida) to facilitate in-person physician consultations. This geographical criterion was essential for standardizing the experimental design’s in-person healthcare professional consultation component and ensuring the ecological validity of the comparative analysis. These criteria were established to ensure participants could meaningfully engage with all experimental conditions while representing the general population encountering medical data interpretation. Exclusion criteria comprised: 1. healthcare professionals or students in healthcare fields; 2. individuals with severe visual or cognitive impairments that would impede experimental task completion; and 3. inability to provide informed consent. Healthcare professionals were excluded to prevent confounding effects of specialized knowledge, consistent with methodological recommendations for patient experience research.
Sample size determination was conducted using a comprehensive a priori power analysis in G*Power 3.1 [65]. Parameters were specified as: medium effect size (f = 0.25) based on meta-analytic findings in comparable health technology studies (Waugaman, 2020; Chen et al., 2021); alpha = 0.05 for statistical significance threshold; power = 0.80 to provide adequate sensitivity for detecting meaningful effects; repeated measures ANOVA with 6 measurements corresponding to the factorial combination of our experimental conditions; estimated correlation among repeated measures = 0.5 based on pilot testing; and accounting for 15 independent variables including primary factors, covariates, and interaction terms. This analysis indicated a minimum sample size of 42 participants, which we increased to 45 to account for potential attrition or unusable data.
The research team implemented a multi-platform recruitment strategy utilizing both digital and traditional approaches to ensure diverse participant representation [66]. Study information, including research objectives, required activities, and estimated time commitment, was disseminated via three primary social media platforms (Instagram, Facebook, and Twitter) using standardized recruitment materials approved by the university’s Institutional Review Board (IRB). Concurrently, recruitment flyers were strategically distributed across the university campus, including the student center, health sciences buildings, main library, and community bulletin boards, with all materials directing interested individuals to contact the Principal Investigator via email. This combined approach generated 109 email inquiries from potential participants, each of which underwent systematic screening against predetermined inclusion and exclusion criteria. Following this comprehensive evaluation process, 45 participants meeting all eligibility requirements were selected, with deliberate attention to maintaining demographic diversity across age, gender, education level, and prior healthcare experience [67]. Selected individuals received electronic informed consent documents via secure email, with participation contingent upon the return of signed consent forms. Upon receipt of the completed consent documentation, participants received formal acceptance notifications that included detailed study participation instructions and compensation information. Table 2 presents the demographic characteristics of the study participants.
Table 2.
Demographic characteristics of study participants (N = 45)
| Characteristic | Category | Frequency | Percentage (%) |
|---|---|---|---|
| Age | 20–30 years | 12 | 26.7 |
| 31–45 years | 15 | 33.3 | |
| 46–60 years | 10 | 22.2 | |
| 61–70 years | 8 | 17.8 | |
| Mean (SD) | 42.3 (15.7) | ||
| Gender | Female | 24 | 53.3 |
| Male | 21 | 46.7 | |
| Race/Ethnicity | Hispanic/Latino | 29 | 64.4 |
| White (Non-Hispanic) | 7 | 15.5 | |
| African American | 5 | 11.2 | |
| Asian | 4 | 8.9 | |
| Occupation | Employed- Full-time | 5 | 11.1 |
| Employed-Part time | 26 | 57.8 | |
| Student | 7 | 15.6 | |
| Retired | 4 | 8.8 | |
| Unemployed | 3 | 6.7 | |
| Education Level | High school | 12 | 26.7 |
| Undergraduate degree | 19 | 42.2 | |
| Graduate degree | 14 | 31.1 | |
| Health Literacy | Low | 10 | 22.2 |
| Moderate | 21 | 46.7 | |
| High | 14 | 31.1 | |
| Prior Technology Experience | Limited | 13 | 28.9 |
| Moderate | 19 | 42.2 | |
| Extensive | 13 | 28.9 | |
| Chronic Health Condition | Yes | 6 | 13.4 |
| No | 39 | 86.6 | |
| Frequency of Medical Data Review | Rarely (<once per year) | 15 | 33.3 |
| Occasionally (1–3 times per year) | 21 | 46.7 | |
| Frequently (>3 times per year) | 9 | 20.0 | |
| Prior Experience with AI Tools | None | 19 | 42.2 |
| Some | 17 | 37.8 | |
| Substantial | 9 | 20.0 |
This study received approval from the IRB at Florida International University (IRB Reference number #115871) and complied with all ethical standards for human subjects research, including the Declaration of Helsinki. Informed Consent and Risk Disclosure: All participants provided written informed consent prior to enrollment. The consent process explicitly addressed the study’s use of Large Language Models and included detailed disclosure of potential risks and limitations. Specifically, participants were informed that: (1) LLMs are not medical professionals and may produce inaccurate, incomplete, or misleading interpretations of medical data; (2) LLM responses should never replace consultation with qualified healthcare professionals; (3) LLMs may make errors in interpreting laboratory values or radiological findings, and participants should not rely on LLM interpretations for medical decision-making; (4) the study examined patient perceptions and experiences, not the clinical accuracy of interpretations; and (5) all medical data used in the study belonged to anonymized cases, not the participants themselves.
The informed consent document emphasized that participation carried no direct clinical risk, as participants were not making actual medical decisions based on the interpretations provided. Data Protection and Confidentiality: All participant data were protected in accordance with IRB-approved protocols and HIPAA regulations. Specifically: (1) all medical data used in the study were completely anonymized and did not belong to participants; (2) all participant responses were assigned study identification numbers rather than names; (3) quantitative data were managed through REDCap secure database with encrypted access; (4) qualitative interviews were audio-recorded, transcribed, and stored in password-protected files; (5) all identifying information was removed from transcripts; and (6) data access was restricted to the research team.
At study completion, participants received a $100 Amazon gift card. Ethical Implications of LLM Errors: Given the healthcare context, particular attention was paid to the ethical implications of potential LLM errors or misinterpretations. The study design addressed these concerns in several ways: (1) participants were explicitly told that medical data did not belong to them, eliminating any possibility of decisions affecting their personal health; (2) all scenarios were presented in an experimental context with clear framing as research participation, not clinical consultation; (3) participants were informed that LLM responses were not validated clinical interpretations; (4) no participant was asked to make or acted upon medical decisions based on the interpretations; and (5) the study measured patient perceptions (comprehension, satisfaction, trust) rather than clinical outcomes.
After completing all six experimental scenarios, participants engaged in a comprehensive debriefing. The debriefing included: (1) explicit restatement that the study examined perceptions of different interpretation methods, not clinical validity; (2) clarification that LLMs should not be used as primary sources for medical information without healthcare professional oversight; (3) discussion of the study’s findings regarding differences between LLM and professional interpretation; (4) discussion of the experimental nature of the data and interpretations; and (5) opportunity for participants to ask questions and discuss their concerns about AI in healthcare. All participants were provided with written information summarizing ethical protections and study resources, in case they had concerns.
This study’s ethical approach reflects the principle that research involving medical AI must include transparent disclosure of AI limitations, particularly regarding potential errors or misinterpretations. While the study examined patient perceptions rather than clinical decision-making, the explicit disclosure of LLM limitations in the informed consent and debriefing processes ensured that participants understood the distinction between research participation and clinical consultation and were fully aware of AI system limitations.
The demographic characteristics of the study sample demonstrate appropriate stratification across key variables. Participants had a mean age of 42.3 years (SD = 15.7), a balanced distribution across age categories, and a slight female predominance (53.3%). The sample reflects Miami’s diverse population, with Hispanic/Latino participants representing the majority (64.4%), followed by White non-Hispanic (15.5%), Black/African American (11.2%), and Asian (8.9%) participants. Most participants were employed part-time (57.8%) or full-time (11.1%), with 15.6% being students. Educational attainment was relatively high, with 73.3% having undergraduate or graduate degrees. Health literacy was predominantly moderate (46.7%) or high (31.1%), and technology experience showed a similar distribution across moderate (42.2%), limited (28.9%), and extensive (28.9%) categories. Only 13.4% reported chronic health conditions, while most participants (46.7%) reported occasional medical data reviews (1–3 times annually). Prior AI tool experience varied: 42.2% reported no experience, underscoring the importance of examining LLM accessibility across different levels of technology familiarity.
Materials
Medical data sets: rationale and characterization
Blood work represents structured, quantitative medical data characterized by: (1) discrete numerical values with well-defined reference ranges; (2) standardized interpretation frameworks with clear decision rules; (3) relatively low visual complexity; and (4) direct comparability across patients. These characteristics create favorable conditions for LLM interpretation, and tasks primarily involve pattern matching and contextual explanation. Medical imaging represents visual, qualitative medical data characterized by: (1) complex visual information requiring spatial reasoning; (2) interpretive judgment necessary to identify findings; (3) inherent interpretive uncertainty; (4) contextual information embedded in visual characteristics; and (5) requirement for simultaneous integration of multiple perspectives. These characteristics create substantially greater interpretive challenges for LLMs, particularly regarding visual understanding and diagnostic uncertainty. This within-study comparison of structured versus visual data modalities addresses a critical research gap. While existing studies examine LLM performance on specific tasks, few systematically compare LLM effectiveness across fundamentally different data types from a patient perspective. By including both modalities, we can determine whether observed differences in patient comprehension and technology acceptance reflect inherent LLM capabilities or artifacts of data structure and complexity. Such findings have substantial practical implications for healthcare decision-making about when LLM augmentation is beneficial.
The experimental stimuli consisted of carefully curated, standardized sets of anonymized medical data developed in consultation with a panel of four board-certified physicians (one general practitioner, one internist, and two radiologists) to ensure clinical relevance and appropriate complexity. Two distinct categories of medical data were created:
The blood work data sets comprised comprehensive laboratory panels, including complete blood count (CBC), comprehensive metabolic panel (CMP), lipid profiles, and thyroid function tests (TFTs). Each panel was designed to include a strategic mixture of values within normal reference ranges (70%) and values outside reference ranges (30%) of varying clinical significance. The abnormal values were further subdivided into three categories of clinical concern: minor (requiring monitoring but no immediate intervention), moderate (suggesting potential pathology requiring follow-up), and significant (indicating clear pathology requiring prompt clinical attention). Values were selected to represent commonly encountered clinical scenarios, including mild anemia, subclinical hypothyroidism, borderline hyperlipidemia, and moderate hyperglycemia. Reference ranges were standardized in accordance with the Clinical and Laboratory Standards Institute (CLSI) guidelines and clearly demarcated on all laboratory reports.
The medical imaging data sets comprised digitized radiological studies, including standardized anteroposterior and lateral knee/elbow X-rays. Images were selected from a teaching archive and anonymized in accordance with HIPAA guidelines, with all identifying information removed and replaced with study identifiers. Each image contained subtle but clinically relevant findings representing three levels of diagnostic complexity: low (e.g., mild degenerative joint changes), moderate (e.g., small osteophyte formation with joint space narrowing), and high (e.g., early signs of inflammatory arthropathy with periarticular osteopenia). Two board-certified radiologists reviewed all images, confirming the presence and classification of findings, and achieved 94% inter-rater agreement on classification. Images were standardized for resolution (2048 × 1536 pixels), contrast, and brightness to ensure consistent presentation across all experimental conditions.
The six experimental scenarios were developed and validated through a rigorous multi-stage process to ensure realism, clarity, and appropriate control of complexity. All medical data were sourced from established clinical teaching archives with full HIPAA anonymization. Blood work values were selected from de-identified laboratory reports representing common clinical scenarios encountered in primary care (mild anemia, subclinical hypothyroidism, borderline hyperlipidemia, moderate hyperglycemia). Medical imaging data were sourced from a radiological teaching archive representing common orthopedic findings. No data belonged to research participants; instead, anonymized clinical cases were randomly assigned to ensure that prior knowledge did not confound. Prior to the main study, all six scenarios underwent pilot testing with 12 participants (stratified by age, education, and technology experience) who did not participate in the main study. Pilot testing assessed: (1) clarity of medical information presentation; (2) appropriateness of complexity level; (3) comprehensibility of instructions; (4) estimated time required for each scenario; (5) technical functionality of the LLM interface; and (6) feasibility of the experimental protocol. Based on pilot feedback, three modifications were implemented: (a) laboratory reference ranges were made more visually prominent on blood work reports; (b) radiological images were standardized to consistent zoom and contrast levels; and (c) instructions for LLM interaction were simplified. Pilot participants confirmed that scenarios felt realistic and clinically plausible.
Complexity was systematically controlled across scenarios through multiple mechanisms. For blood work, abnormal findings were categorized into three complexity levels (minor, moderate, significant) and distributed evenly across the six scenarios, with each participant encountering one finding at each complexity level; abnormal values ranged from 30 to 40% across scenarios to maintain consistency; and terminology and reference ranges were standardized using UMLS. For medical imaging, findings were classified into three complexity levels (low, moderate, high), with images randomly assigned to scenarios to ensure balanced complexity distribution; inter-rater agreement between two radiologists confirmed 94% concordance in complexity classification; and image quality, resolution, and viewing parameters were standardized across all radiographs.
To ensure scenarios represented clinically realistic situations, all medical data were reviewed and approved by the physician panel (one general practitioner, one internist, two radiologists) who confirmed: (1) the clinical plausibility of all values and findings; (2) that scenarios represented commonly encountered clinical presentations; (3) that the mixture of normal and abnormal findings was representative of typical clinical practice; and (4) that complexity levels appropriately reflected the difficulty of real-world interpretation. The physician panel specifically verified that none of the abnormal findings represented rare or unusual clinical presentations unlikely to occur in actual clinical practice. Medical data presentation was standardized across all three interpretation modalities (phone consultation, in-person consultation, LLM interaction) to ensure that differences in outcomes reflected differences in interpretation method rather than differences in data presentation. Specifically: (1) the same anonymized medical data were used across all modalities for each participant; (2) information content was held constant (e.g., physicians provided all relevant values in the same order as presented in the written report); (3) standard medical terminology was used across all modalities using UMLS definitions; and (4) complexity level was maintained equivalent across modalities.
To ensure experimental control while maintaining ecological validity, all participants were presented with identical standardized medical data sets across the six interpretation scenarios, with the order of presentation randomized and counterbalanced using a Latin square design to mitigate potential order and learning effects. Importantly, the medical data used in this study did not belong to the participants themselves; instead, anonymized clinical cases were randomly assigned to each participant to eliminate the confounding effect of prior knowledge or preconceived interpretations that might occur with personal health data [68]. This methodological approach ensured that participants’ responses were based solely on the information presented through the various interpretation modalities, rather than on their prior familiarity with their medical results. Medical terminology was standardized across all materials using preferred terms from the Unified Medical Language System (UMLS) to ensure consistency in communication regardless of interpretation modality, further supporting the validity of cross-condition comparisons [69].
LLM configuration
We employed ChatGPT-4o, OpenAI’s most advanced general-purpose large language model at the time of the study’s implementation (March and April 2025), for the LLM interaction scenarios. This selection was justified through a systematic assessment of contemporary commercially available LLMs, with GPT-4o demonstrating superior performance on healthcare benchmarks, including medical knowledge retrieval, explanation accuracy, and natural language understanding of patient queries [70]. While not specifically fine-tuned on medical datasets, recent evaluations have documented GPT-4o’s robust capabilities in addressing healthcare-related queries, with performance approaching that of specialized medical models on multiple benchmarks [71].
Moreover, our deliberate use of ChatGPT-4o as the LLM representative in this study reflects our research focus on accessibility and ecological validity. We specifically chose a widely available general-purpose model rather than a specialized medical AI system to investigate whether the general public would use accessible LLMs for medical data interpretation rather than seeking domain-specific models with limited availability [72]. While this choice enhances the real-world applicability of our findings, it also means our results may differ from those observed with highly specialized medical LLMs. GPT-4o’s multimodal capabilities were particularly advantageous for our study design, enabling the model to process visual information from radiological images alongside textual queries, a critical feature for the medical imaging interpretation scenarios [73]. Multiple studies have demonstrated that patients increasingly turn to general-purpose LLMs like ChatGPT for interpreting medical information, underscoring the ecological validity of evaluating this widely accessible tool rather than specialized models with limited public availability [74]. Additionally, GPT-4o’s retrieval-augmented generation capabilities provide access to a substantial knowledge base of medical literature and clinical guidelines, enabling accurate interpretation of laboratory values and radiological findings without requiring domain-specific fine-tuning [75, 76].
To optimize performance for medical data interpretation tasks, we configured a CustomGPT (named “Medical Explainer AI”) using GPT-4o with a specialized system prompt: “You are a medical assistant providing clear, accurate explanations of medical data, including blood test results and radiological images. Explain findings in plain language, highlight abnormal values, provide context for their significance, and answer questions about medical terminology. Always include reference ranges for lab values. Acknowledge your limitations, avoid definitive diagnoses, and recommend consulting healthcare professionals for medical decisions. Strive for a balance between accuracy and accessibility in explanations.” This system prompt was developed through iterative refinement based on clinical communication guidelines and patient health literacy research [77].
To mitigate potential interaction biases arising from varying prompt-engineering capabilities among participants, standardized instructional protocols were developed and provided to all participants. These instructions included: (1) sample starter prompts for initiating interpretation sessions (e.g., “Please explain my blood test results, focusing on any abnormal values” or “What does this X-ray show about my knee?”); (2) follow-up probing templates for seeking clarification or additional information (e.g., “Can you explain [specific term] in simpler language?” or “What is the clinical significance of [specific finding]?”); and (3) comparative question formats for understanding result implications (e.g., “How do these results compare to normal ranges?” or “What might have caused these findings?”). These standardized prompts were developed through iterative testing with patient focus groups to ensure comprehensibility while maintaining consistent interaction patterns across participants [78].
The GPT-4o interface was accessed through a secured CustomGPT (i.e., Medical Explainer AI) link with a standardized layout and functionality across all experimental sessions, ensuring a uniform user experience. System response time, interaction history, and query patterns were logged for subsequent analysis of interaction characteristics and their relationship to outcome variables. By using this consistently configured LLM via a controlled-access mechanism, our study achieved a balance between ecological validity and experimental control, thereby enhancing the reliability and precision of our comparative analysis.
Prompt structure and interface design
To enhance reproducibility and interpretability, detailed specifications of the LLM interaction structure are provided below.
System Prompt: The CustomGPT was configured with the following system prompt: “You are a medical assistant providing clear, accurate explanations of medical data, including blood test results and radiological images. Explain findings in plain language, highlight abnormal values, provide context for their significance, and answer questions about medical terminology. Always include reference ranges for lab values. Acknowledge your limitations, avoid definitive diagnoses, and recommend consulting healthcare professionals for medical decisions. Strive for a balance between accuracy and accessibility in explanations.”
Interaction Structure - Guided but Open-Ended: The study employed a semi-structured interaction approach, guided by standardized starter prompts while allowing full openness for follow-up questions. Specifically, participants received: (1) standardized initial prompt templates (e.g., “Please explain my blood test results focusing on any abnormal values” or “What does this X-ray show about my knee?”); (2) follow-up probing templates for clarification (e.g., “Can you explain [specific term] in simpler language?” or “What is the clinical significance of [specific finding]?”); and (3) comparative question formats (e.g., “How do these results compare to normal ranges?” or “What might have caused these findings?”). After receiving these standardized starter prompts, participants were explicitly instructed that they could ask ANY additional questions they wished in plain language, with no restrictions on follow-up inquiries. This semi-structured approach ensured: (a) comparable initial interaction across participants, enabling valid outcome comparison; (b) open-ended capability for participants to explore information according to individual needs, supporting perceived control outcomes; and (c) natural conversational flow rather than scripted rigidity.
User Interface Design: CustomGPT was accessed via a secure web link, providing a standardized interface consistent across all participant sessions. Interface specifications included: (1) clean, uncluttered text-based chat interface with minimal visual distractions; (2) medical data (blood test results or radiological images) displayed prominently at the top of the interface with the chat conversation below; (3) uniform font, text size, and color scheme across all sessions; (4) clear visual separation between participant queries and LLM responses; (5) response timestamp logging for session analysis; (6) ability to scroll through conversation history; and (7) no additional features, plugins, or recommendations that might influence interpretation. The interface was intentionally designed to be neutral and unbiased, minimizing visual cues that might subtly influence trust or perceived authority compared to those in healthcare professional consultations.
Potential Influence on Outcomes: These design elements likely influenced measured outcomes in several ways. The semi-structured approach (standardized starters + open-ended follow-up) may have enhanced perceived control by allowing participants to direct their own questioning while providing enough structure to ensure comparable baseline comprehension. The explicit instruction that follow-up questions could be asked in any format likely increased perceived control relative to time-constrained professional consultations. The plain-language system prompt emphasizing accessibility may have enhanced comprehension compared to more technical language. The interface’s neutral design may have reduced experimenter-induced biases in trust formation compared with phone consultations that show a physician’s face or hear a physician’s tone of voice. Conversely, the text-based interface lacked nonverbal cues and human presence, which typically enhance trust and confidence in healthcare settings. The standardized starter prompts ensured comparable initial comprehension but may have reduced some participants’ sense of autonomy compared to fully unrestricted question-asking. These design choices represent necessary tradeoffs between experimental control (enabling valid comparison across conditions) and ecological validity (resembling real-world LLM use patterns).
Reproducibility Specifications: To enable reproducibility, the following specifications are provided: (1) LLM model: GPT-4o (OpenAI); (2) CustomGPT configuration: Medical Explainer AI; (3) system prompt: [provided above]; (4) standardized starter prompts: [see Appendix A]; (5) interface platform: OpenAI CustomGPT web interface; (6) session duration limit: 30 minutes maximum; (7) no plugins, web access, or external integrations; (8) response temperature setting: 0.7 (balancing consistency and natural variation); (9) interaction logs preserved with timestamping for subsequent analysis.
Healthcare professional protocols
Four board-certified physicians were strategically recruited for the healthcare professional consultation scenarios to conduct the interpretation sessions: two primary care physicians (PCPs) for blood work interpretation and two radiologists for medical imaging interpretation. This specialized allocation was deliberately designed to control for potential familiarity effects that might influence participant perceptions across different communication modalities [79, 80]. Specifically, one PCP was consistently assigned to blood work interpretation via telephone consultations, while the second PCP exclusively conducted in-person blood work consultations. Similarly, one radiologist was dedicated to telephone-based medical imaging interpretation, and the second radiologist performed only in-person imaging consultations.
This methodological approach ensured that participants experienced distinct healthcare professionals across different modalities, thereby eliminating potential carryover effects or prior rapport that might confound comparisons between phone and in-person consultations [81]. The participating physicians possessed comparable clinical qualifications (mean clinical experience = 11.8 years, SD = 3.2) and underwent identical standardized training to minimize provider-specific variations in communication style and content delivery [82]. All four physicians completed a comprehensive 2-hour-and-a-half online training program developed specifically for this study. The training encompassed: 1. familiarization with the standardized medical data sets; 2. implementation of structured communication protocols; 3. calibration exercises to ensure consistency in explanation depth and terminology usage; and 4. simulated interpretation sessions with standardized patients. This intensive preparation established a foundation for consistent patient-provider interactions while preserving naturalistic communication patterns [83].
The communication protocols were developed through a modified Delphi process involving healthcare communication experts, physicians, and patient representatives to establish best practices for medical data interpretation while maintaining a naturalistic interaction style. The protocols specified a three-phase structure for each consultation: introduction and rapport-building, systematic data presentation and interpretation, and summary with opportunity for questions. For abnormal findings, the protocol mandated an explanation of: the nature of the abnormality, its potential clinical significance, appropriate contextual information, and general follow-up considerations without providing personalized recommendations. To ensure standardization while preserving ecological validity, physicians were provided with communication guidance specifying key points for each data element but were instructed to maintain their natural communication style. Fidelity to the protocol was ensured through pilot testing with standardized patients and assessment using the Medical Communication Competence Scale [84], with all physicians achieving competency scores exceeding 85%. Physicians were informed of the general research focus on medical data interpretation, but were blinded to specific hypotheses regarding the comparison with LLM interaction to prevent potential biases in the communication approach.
Experiment procedure
The experimental protocol followed a systematic, multi-phase process conducted during March and April 2025, designed to maximize internal validity while ensuring ecological validity of participants’ experiences. Prior to enrollment, all participants completed a comprehensive screening to verify eligibility according to the inclusion/exclusion criteria. The experimental phase employed a within-subjects repeated-measures design, in which each participant (N = 45) experienced all six interpretation scenarios (3 communication modalities × 2 medical data types), with the order randomized independently for each participant using a computer-generated sequence to mitigate systematic order effects. To minimize potential carryover effects while maintaining practical feasibility for participant retention, a mandatory minimum washout period of 3 days between scenario sessions was implemented, based on cognitive interference research demonstrating significant decay in specific medical information recall after this interval [85].
Participants first received standardized instructions for each experimental session regarding the scenario they would experience. Participants received a comprehensive laboratory report in a standard clinical format, with clear reference ranges for interpreting blood work. Participants were presented with high-resolution digital radiographic images on calibrated medical-grade monitors with standardized viewing conditions for medical imaging interpretation. Also, the blood work report and medical imaging were provided in digital format, along with instructions on how to upload them to CustomGPT. The four board-certified physicians (two primary care physicians for blood work and two radiologists for imaging) agreed to dedicate 7 days across different weeks to study participation, creating multiple appointment slots to accommodate participants’ scheduling needs.
For LLM conditions, participants accessed the CustomGPT interface (Medical Explainer AI) via a standardized, secure link, where they could upload the assigned blood test results or radiographic images and follow structured instructions. These sessions were conducted at participants’ preferred locations using devices that met minimum technical specifications. For healthcare professional phone consultations, participants received calls at pre-arranged convenient times while in their home environments, with their assigned physician (either a primary care physician for blood work or a radiologist for imaging) conducting the consultation via an institutional telehealth platform. For in-person consultations, participants visited the respective physicians’ clinical facilities during after-hours timeframes specifically allocated for study purposes, with appointments scheduled to accommodate both participant and physician availability. Each interpretation session followed a standardized procedure wherein participants had up to 30 minutes to interact with the assigned interpretation method. For example, a participant randomly assigned to first evaluate blood test results using the LLM interface completed that session, then, after a minimum three-day interval, engaged in a phone consultation with the designated primary care physician to discuss blood test results, followed by an in-person consultation with the second primary care physician after another minimum three-day interval. This pattern continued for the medical imaging interpretation scenarios involving the two radiologists, with a minimum three-day washout period between all sessions.
All sessions involving a doctor were audio-recorded, and, for LLM conditions, screen captures were also recorded with participant consent. Immediately after each interpretation session, participants completed a comprehensive battery of post-scenario assessments that measured all dependent variables. These assessments were administered via a secure electronic survey platform, and measures were presented in random order to prevent systematic response bias. After completing quantitative measures in each experiment, participants engaged in a 15-minute semi-structured interview exploring their subjective experiences, guided by a standardized interview guide. Moreover, after completing all six experimental conditions, participants completed an online integrative-comparative questionnaire that required explicit reflection on their experiences across all interpretation methods. This instrument incorporated both quantitative preference ratings and open-ended qualitative responses regarding the perceived advantages and limitations of each modality. Based on a preliminary analysis of experimental data and purposive sampling to ensure diversity in demographics and response patterns, a subset of participants (n = 15) was invited to comprehensive follow-up interviews (45–60 minutes) conducted by the principal investigator, who used grounded theory-informed qualitative analysis.
Throughout all phases, rigorous quality control procedures were implemented, including standardized protocol checklists, regular calibration meetings for the investigator team, and independent verification of protocol adherence through random session audits. All data were securely stored in accordance with IRB-approved procedures, with quantitative data managed in the REDCap secure database and qualitative data processed using NVivo 14. Interaction logs from LLM sessions were preserved with detailed timestamping to enable analysis of conversation patterns, response latencies, and question typologies.
Measures and measurement model assessment
Operationalization of patient empowerment
Within the context of this study, patient empowerment is operationalized as a multidimensional construct encompassing patients’ psychological and behavioral capacities to engage actively in health information processing and decision-making. Specifically, we define patient empowerment through three measurable dimensions (i.e., perceived control, decision confidence, and health self-efficacy). Perceived control (measured via the Health Care Empowerment Inventory): Reflects the extent to which patients feel they have agency and autonomy in understanding and managing their medical information. Higher scores (ranging from 1 to 7) indicate greater perceived control over the interpretation process and their health information. A score of 7 represents maximum perceived control, while a score of 1 indicates minimal control. Decision confidence (measured via the Decisional Conflict Scale, reverse-coded): Captures patients’ confidence in understanding medical findings and in their capacity to make informed decisions based on the information interpreted. Higher scores (1–7) represent greater confidence in decision-making capability and reduced decisional conflict. A score of 7 indicates high decision confidence; a score of 1 indicates low confidence with substantial decision ambiguity. Health self-efficacy (measured via the Patient Activation Measure): Assesses patients’ self-perceived ability to understand, manage, and act on their health information. Higher scores (1–7) indicate stronger self-efficacy beliefs about one’s capacity to comprehend and engage with medical data. A score of 7 represents high health self-efficacy; a score of 1 represents low self-efficacy.
For all patient empowerment dimensions, higher scores indicate more positive empowerment outcomes, greater perceived control, higher decision confidence, and stronger self-efficacy beliefs. This unidirectional scale prevents interpretive confusion and aligns with standard psychometric conventions in health services research. Patient empowerment across these three dimensions reflects the extent to which an interpretation method enables patients to feel capable, informed, and agentic in their engagement with medical information.
Study measures
Based on the integrated theoretical framework, the variables shown in Table 3 were measured using validated instruments. All multi-item scales were assessed for reliability using Cronbach’s alpha, and validity was confirmed through confirmatory factor analysis.
Table 3.
Study measures and instruments
| Construct Domain | Measure | Instrument | Source | Description | Scale Direction |
|---|---|---|---|---|---|
| Information Processing Variables | |||||
| Diagnostic Comprehension | 10-item objective test | Adapted from Medical Data Interpretation Test | Murphy, Davis [86] | Assesses understanding of key elements in presented medical data through multiple-choice and short-answer questions | Higher = Better comprehension |
| Information Satisfaction | Medical Information Satisfaction Scale (MISS-21) | 21-item questionnaire | Meakin and Weinman [87] | Measures satisfaction with clarity, completeness, and usefulness of medical information provided | Higher = Greater satisfaction |
| Cognitive Load | NASA Task Load Index (NASA-TLX) | 6-dimensional workload assessment | Hart [88] | Evaluates mental, physical, and temporal demands, effort, performance, and frustration levels during interpretation | Lower = Reduced cognitive load |
| Patient Empowerment Variables | |||||
| Perceived Control | Health Care Empowerment Inventory (HCEI) | 8-item scale | Johnson, Rose [89] | Measures perceived ability to participate actively in healthcare decisions and manage health information | Higher = Greater perceived control |
| Decision Confidence | Decision Self-Efficacy Scale (DSES) | 11-item questionnaire | O’Connor [90] | Assesses confidence in making informed medical decisions based on interpreted information | Higher = Greater decision confidence |
| Health Self-Efficacy | Patient Activation Measure (PAM-13) | 13-item scale | Hibbard, Mahoney [91] | Evaluates knowledge, skills, and confidence for self-management of health and healthcare | Higher = Greater self-efficacy |
| Technology Acceptance Variables | |||||
| Performance Expectancy | UTAUT Scale | 4-item adapted scale | Venkatesh, Morris [11] | Measures the degree to which an individual believes that using the interpretation method will help them understand their medical data better | Higher = Higher performance expectancy |
| Effort Expectancy | UTAUT Scale | 4-item adapted scale | Venkatesh, Morris [11] | Assesses the degree of ease associated with using and understanding the interpretation method | Higher = Perceived ease of use |
| Behavioral Intention to Use | UTAUT Scale | 3-item adapted scale | Venkatesh, Morris [11] | Evaluates likelihood of using the interpretation method in the future | Higher = Stronger intention to use |
| Additional Outcome Variables | |||||
| Anxiety | State-Trait Anxiety Inventory (STAI) | 20-item state anxiety subscale |
Spielberger, Gonzalez-Reigosa [92]; Kvaal, Ulstein [93] |
Measures current levels of anxiety experienced during medical data interpretation | Lower = Reduced anxiety |
| Perceived Accuracy | Perceived Information Accuracy Assessment (PIAA) | 5-item scale | Appelman and Sundar [94] | Evaluates perceived correctness, credibility, and reliability of the interpreted medical information | Higher = Higher perceived accuracy |
| Trust in Interpretation | Trust in Medical Technology Scale | 10-item adapted scale | Montague [95] | Assesses trust in the interpretation method’s reliability, competence, and dependability | Higher = Greater trust |
| Time Efficiency | Interpretation Session Duration (Minutes) | Objective timestamp measurement | N/A | Records the total time required to complete the interpretation session from start to finish | Lower = Better efficiency |
Prior to testing the hypothesized relationships, we conducted a comprehensive assessment of all measurement instruments. Internal consistency reliability was evaluated using Cronbach’s alpha, with all multi-item scales demonstrating good to excellent reliability. The Information Processing variables showed strong reliability coefficients: Diagnostic Comprehension (α = 0.87), Information Satisfaction (α = 0.92), and Cognitive Load (α = 0.84). Patient Empowerment measures similarly demonstrated robust internal consistency: Perceived Control (α = 0.89), Decision Confidence (α = 0.91), and Health Self-Efficacy (α = 0.88). Technology Acceptance variables maintained excellent reliability: Performance Expectancy (α = 0.94), Effort Expectancy (α = 0.92), and Behavioral Intention to Use (α = 0.95). Additional outcome measures also showed strong internal consistency: Anxiety (α = 0.90), Perceived Accuracy (α = 0.88), and Trust in Interpretation (α = 0.93).
Confirmatory factor analysis (CFA) was conducted to verify the construct validity of our measurement model. The initial CFA model demonstrated acceptable but suboptimal fit (χ2 (1423) = 2876.42, p < 0.001; CFI = 0.891; TLI = 0.883; RMSEA = 0.072 [90% CI: 0.068, 0.076]; SRMR = 0.064). Based on modification indices, we identified four problematic items with significant cross-loadings or low factor loadings (<0.50). After removing these items (one each from the Information Satisfaction, Cognitive Load, Perceived Control, and Trust scales), the revised measurement model demonstrated a good fit (χ2 (1214) = 1986.35, p < 0.001; CFI = 0.942; TLI = 0.936; RMSEA = 0.059 [90% CI: 0.054, 0.064]; SRMR = 0.048). All items loaded significantly on their respective factors (standardized loadings ranging from 0.68 to 0.94, p < 0.001), supporting convergent validity. The average variance extracted (AVE) for each construct exceeded 0.50 (range: 0.61–0.83), and the square root of each construct’s AVE was greater than its correlation with other constructs, establishing discriminant validity. These results indicate that our measurement instruments demonstrated satisfactory psychometric properties for testing the structural relationships in our theoretical model.
Data analysis
Quantitative analysis
The quantitative data analysis employed a comprehensive, multi-method approach to examine the complex relationships between interpretation methods and outcome variables. Primary analyses were conducted using IBM SPSS Statistics 28.0 and AMOS 28.0 structural equation modeling software. To address our main research questions regarding differences across interpretation scenarios, we conducted a repeated-measures analysis of variance (ANOVA) with interpretation method (healthcare professional via phone, healthcare professional in person, and LLM interaction) and data type (blood work results and medical imaging) as within-subjects factors. This approach allowed us to examine the main effects of each factor and their interactions with all dependent variables. Post hoc pairwise comparisons with a Bonferroni correction for multiple comparisons were conducted to identify specific differences between interpretation methods while controlling for Type I error rates.
We developed and evaluated a structural equation model (SEM) using maximum likelihood estimation to test our integrated theoretical framework linking interpretation methods to information processing, patient empowerment, and technology acceptance constructs. The model specified direct paths from the interpretation method to the information processing variables, from the information processing variables to the patient empowerment variables, and from the information processing and patient empowerment variables to the technology acceptance variables. Model fit was assessed using multiple indices, including chi-square goodness-of-fit, comparative fit index (CFI), Tucker-Lewis index (TLI), root mean square error of approximation (RMSEA), and standardized root mean square residual (SRMR). Bootstrap resampling procedures with 5,000 samples were employed to test direct and indirect effects and generate bias-corrected confidence intervals.
To identify key predictors of outcome variables across different interpretation scenarios, we conducted multiple regression analyses with hierarchical entry of predictor blocks. Demographic variables were entered in the first block, information processing variables in the second, and patient empowerment variables in the third. This approach allowed us to determine the unique contribution of each variable set while controlling for variables already entered. Moderation analyses were performed using the PROCESS macro for SPSS to examine whether participant characteristics (age, gender, education level, and prior health literacy) moderated the relationship between the interpretation method and outcome variables. Moderation models were tested with 5,000 bootstrap samples and heteroscedasticity-consistent standard errors. Finally, mediation analyses were conducted to test whether information-processing and patient-empowerment variables mediated the effect of the interpretation method on technology acceptance. We tested both simple mediation models (examining each mediator independently) and serial mediation models (examining sequential pathways through multiple mediators) using the bootstrapping approach recommended by Hayes [96]. These analyses allowed us to decompose the total effect of the interpretation method into direct and indirect components, providing insights into the psychological mechanisms through which different interpretation methods influence technology acceptance outcomes. All statistical tests employed a significance threshold of α = 0.05, with exact p-values reported where appropriate.
Qualitative analysis
The qualitative data collection in this study followed a systematic, multi-phase approach integrated with the experimental design. Following each interpretation scenario, participants engaged in brief semi-structured interviews lasting approximately 15 minutes to capture their immediate reactions and experiences. These post-scenario interviews focused on perceptions of information clarity, emotional responses, and comparative reflections on previously experienced scenarios. All participants (N = 45) completed these interviews after each of the six scenarios, resulting in 270 brief, scenario-specific interview sessions. Interview questions were developed based on the literature to ensure relevance and comprehensiveness. Sample prompts included: “How would you describe your experience interpreting your medical results through this method?” and “What aspects of this interpretation method were most helpful or challenging for you?” The interview questions and guidelines are outlined in Appendix B.
After completing all six scenarios, a purposively selected subset of participants (n = 15) underwent in-depth follow-up interviews lasting 45–60 minutes. This subset was strategically selected to ensure representation across demographic variables, including age, gender, education level, and baseline health literacy, with particular attention to participants who exhibited varying response patterns across the interpretation methods. These comprehensive interviews explored broader themes and allowed participants to reflect on their complete experience across all scenarios. A semi-structured interview guide organized around the theoretical constructs of information processing, patient empowerment, and technology acceptance guided these conversations while allowing flexibility to explore emergent topics. To minimize bias, all interviews were conducted by a researcher with a background in patient-facing communications who was not involved in the quantitative data collection.
All interview sessions were audio-recorded with participant consent and professionally transcribed. Field notes documenting non-verbal cues, emotional reactions, and environmental factors supplemented the audio recordings. For LLM interaction scenarios, complete chat logs were preserved and integrated with interview data to provide a contextual understanding of participant experiences. Transcripts were validated for accuracy by comparing them with the original recordings before being imported into NVivo 14 for organization and analysis. This study employed open, axial, and selective coding techniques, as outlined by Corbin and Strauss (1990), to identify conceptually similar themes expressed by interview participants. Open coding involves identifying and defining key concepts, properties, and dimensions analytically. Axial coding then established relationships among these categories by linking them to the open codes based on conceptual similarities. Finally, selective coding integrated and refined these categories into higher-level theoretical constructs, offering a cohesive explanation of the concerns and benefits of each interaction scenario. This comprehensive qualitative data collection approach yielded rich, contextually grounded insights that complemented the quantitative metrics and allowed a more detailed understanding of participant experiences across the different interpretation modalities.
Results
Descriptive statistics
Values in Table 4 represent means (SD). All psychological scales range from 1 to 7, where higher values indicate more favorable outcomes. For example, greater comprehension, satisfaction, empowerment, acceptance, trust, and lower cognitive load or anxiety (after reverse-coding where applicable). Diagnostic Comprehension is scored 0–10 (correct answers), and Time Efficiency is measured in minutes (lower values = faster completion). Reliability coefficients (Cronbach’s alpha) for all multi-item scales exceeded 0.80.
Table 4.
Descriptive statistics for dependent variables across six experimental scenarios
| Variable | Scale & Direction | Blood Work + Phone | Blood Work + In-Person | Blood Work + LLM | Imaging + Phone | Imaging + In-Person | Imaging + LLM |
|---|---|---|---|---|---|---|---|
| Information Processing Variables | |||||||
| Diagnostic Comprehension | Understanding of Medical Data (0–10 scale; higher = better comprehension) | 6.8 (1.3) | 7.9 (1.2) | 8.3 (1.0) | 6.2 (1.5) | 7.3 (1.4) | 7.3 (1.3) |
| Information Satisfaction | Satisfaction with Clarity & Completeness of Information Provided (1–7 scale; higher = more satisfied) | 4.9 (1.1) | 5.9 (0.8) | 5.5 (1.0) | 4.7 (1.2) | 5.7 (0.9) | 5.1 (1.1) |
| Cognitive Load | Mental Effort Required to Process Information (1–7 scale; lower = less cognitive burden) | 4.4 (1.1) | 3.7 (1.0) | 3.0 (0.9) | 4.6 (1.3) | 3.9 (1.1) | 3.4 (1.0) |
| Patient Empowerment Variables | |||||||
| Perceived Control | Ability to Participate in Healthcare Decisions (1–7 scale; higher = greater control) | 4.3 (0.9) | 5.0 (0.8) | 5.8 (0.7) | 4.1 (1.0) | 4.8 (0.9) | 5.4 (0.8) |
| Decision Confidence | Confidence in Making Informed Medical Decisions (1–7 scale; higher = more confident) | 4.7 (1.0) | 5.9 (0.7) | 5.2 (0.9) | 4.3 (1.1) | 5.5 (0.8) | 4.6 (1.0) |
| Health Self-Efficacy | Knowledge, Skills & Confidence for Self-Management (1–7 scale; higher = stronger capability) | 4.7 (1.0) | 5.5 (0.9) | 5.6 (0.8) | 4.5 (1.1) | 5.1 (1.0) | 5.2 (0.9) |
| Technology Acceptance Variables | |||||||
| Performance Expectancy | Belief That Method Will Help Understand Medical Data (1–7 scale; higher = greater benefit expected) | 5.0 (1.0) | 5.6 (0.9) | 6.1 (0.8) | 4.8 (1.1) | 5.4 (1.0) | 5.5 (0.9) |
| Effort Expectancy | Ease of Using the Interpretation Method (1–7 scale; higher = easier to use) | 4.9 (1.0) | 5.4 (0.9) | 6.3 (0.7) | 4.7 (1.1) | 5.2 (1.0) | 5.9 (0.8) |
| Behavioral Intention to Use | Likelihood of Using Method in Future (1–7 scale; higher = greater intention to use) | 4.6 (1.2) | 5.5 (1.0) | 6.0 (0.9) | 4.4 (1.3) | 5.3 (1.1) | 5.4 (1.0) |
| Additional Outcome Variables | |||||||
| Anxiety | Emotional Distress Experienced During Interpretation (1–7 scale; lower = less anxiety) | 3.5 (1.0) | 2.2 (0.7) | 2.8 (0.9) | 3.7 (1.1) | 2.4 (0.8) | 3.0 (1.0) |
| Perceived Accuracy | Belief That Information is Correct & Reliable (1–7 scale; higher = more accurate) | 5.6 (0.8) | 6.3 (0.6) | 5.4 (0.9) | 5.4 (0.9) | 6.1 (0.7) | 5.0 (1.1) |
| Trust in Interpretation | Confidence in Reliability & Competence of Method (1–7 scale; higher = more trustworthy) | 5.4 (0.9) | 6.1 (0.7) | 4.8 (1.1) | 5.2 (1.0) | 5.9 (0.8) | 4.4 (1.2) |
| Time Efficiency | Duration of Interpretation Session (minutes; lower = more efficient) | 17.9 (5.8) | 23.8 (6.9) | 11.7 (3.9) | 19.5 (6.3) | 25.2 (7.6) | 12.9 (4.4) |
Note: Values represent means with standard deviations in parentheses. Each row includes a description of what the variable measures and the scale direction (higher = better or lower = better). All scale ranges and their meanings are indicated in the variable descriptions
The three patient empowerment variables (i.e., perceived control, decision confidence, and health self-efficacy) were measured using 1–7 Likert-type scales, with higher scores indicating greater empowerment outcomes. Thus, a score of 7 represents maximum empowerment, while a score of 1 represents minimal empowerment across all three dimensions. Participants in LLM interaction demonstrated notably higher perceived control scores (M = 5.6, SD = 0.8) than in phone consultation (M = 4.2, SD = 1.0), reflecting the self-directed nature of AI-mediated interactions, in which patients control the pace, depth, and direction of information exploration. Decision confidence and health self-efficacy were similarly higher in the in-person and LLM conditions than in the phone consultation condition, indicating that both face-to-face interaction and autonomous, technology-mediated interpretation enhance patients’ confidence in understanding medical information and their capacity to engage with health decisions.
Hypothesis testing results
Information processing variables
As shown in Table 5, repeated measures ANOVA revealed a significant main effect of the interpretation method on diagnostic comprehension, F(2, 88) = 18.72, p < 0.001, η2 = 0.32. Post-hoc comparisons indicated that LLM interaction (M = 7.8, SD = 1.2) resulted in significantly higher comprehension scores than healthcare professional phone consultation (M = 6.5, SD = 1.4, p < 0.001) and was comparable to in-person consultation (M = 7.6, SD = 1.3, p = 0.42). There was also a significant main effect of data type, F(1, 44) = 12.35, p < 0.001, η2 = 0.22, with blood work interpretation (M = 7.5, SD = 1.3) yielding higher comprehension than medical imaging (M = 6.9, SD = 1.5). A significant interaction effect, F(2, 88) = 7.63, p < 0.01, η2 = 0.15, indicated that the benefit of LLM was greater for blood work than for imaging interpretation.
Table 5.
ANOVA results for information processing variables
| Variable | Effect | F | df | p | η2 |
|---|---|---|---|---|---|
| Diagnostic Comprehension | Interpretation Method | 18.72 | (2, 88) | <0.001 | 0.32 |
| Data Type | 12.35 | (1, 44) | <0.001 | 0.22 | |
| Interpretation Method × Data Type | 7.63 | (2, 88) | <0.01 | 0.15 | |
| Information Satisfaction | Interpretation Method | 9.85 | (2, 88) | <0.001 | 0.18 |
| Data Type | 2.43 | (1, 44) | 0.126 | 0.05 | |
| Interpretation Method × Data Type | 1.87 | (2, 88) | 0.160 | 0.04 | |
| Cognitive Load | Interpretation Method | 14.31 | (2, 88) | <0.001 | 0.24 |
| Data Type | 3.27 | (1, 44) | 0.077 | 0.07 | |
| Interpretation Method × Data Type | 0.92 | (2, 88) | 0.402 | 0.02 |
For information satisfaction, a significant main effect of the interpretation method was found, F(2, 88) = 9.85, p < 0.001, η2 = 0.18. In-person consultation (M = 5.8, SD = 0.9) produced the highest satisfaction, followed by LLM interaction (M = 5.3, SD = 1.1) and phone consultation (M = 4.8, SD = 1.2). All pairwise comparisons were significant (p < 0.05). Analysis of cognitive load revealed a significant main effect of the interpretation method, F(2, 88) = 14.31, p < 0.001, η2 = 0.24. LLM interaction (M = 3.2, SD = 1.0) resulted in significantly lower cognitive load than both phone consultation (M = 4.5, SD = 1.2, p < 0.001) and in-person consultation (M = 3.8, SD = 1.1, p < 0.05).
Patient empowerment variables
As shown in Table 6, a significant main effect of the interpretation method was observed for perceived control, F(2, 88) = 21.45, p < 0.001, η2 = 0.33. LLM interaction (M = 5.6, SD = 0.8) resulted in significantly higher perceived control than both phone consultation (M = 4.2, SD = 1.0, p < 0.001) and in-person consultation (M = 4.9, SD = 0.9, p < 0.01). For decision confidence, results showed a significant main effect of interpretation method, F(2, 88) = 16.78, p < 0.001, η2 = 0.28. In-person consultation (M = 5.7, SD = 0.8) yielded the highest decision confidence, followed by LLM interaction (M = 4.9, SD = 1.0) and phone consultation (M = 4.5, SD = 1.1). Data type also showed a significant main effect, F(1, 44) = 9.25, p < 0.01, η2 = 0.17, with greater confidence for blood work (M = 5.3, SD = 0.9) than imaging (M = 4.8, SD = 1.1). For health self-efficacy, a significant main effect of the interpretation method was found, F(2, 88) = 11.24, p < 0.001, η2 = 0.20. Both LLM interaction (M = 5.4, SD = 0.9) and in-person consultation (M = 5.3, SD = 1.0) resulted in significantly higher self-efficacy than phone consultation (M = 4.6, SD = 1.1, p < 0.001), but did not differ significantly from each other (p = 0.61).
Table 6.
ANOVA results for patient empowerment variables
| Variable | Effect | F | df | p | η2 |
|---|---|---|---|---|---|
| Perceived Control | Interpretation Method | 21.45 | (2, 88) | <0.001 | 0.33 |
| Data Type | 3.63 | (1, 44) | 0.063 | 0.08 | |
| Interpretation Method × Data Type | 1.39 | (2, 88) | 0.254 | 0.03 | |
| Decision Confidence | Interpretation Method | 16.78 | (2, 88) | <0.001 | 0.28 |
| Data Type | 9.25 | (1, 44) | <0.01 | 0.17 | |
| Interpretation Method × Data Type | 2.43 | (2, 88) | 0.094 | 0.05 | |
| Health Self-Efficacy | Interpretation Method | 11.24 | (2, 88) | <0.001 | 0.20 |
| Data Type | 4.11 | (1, 44) | 0.049 | 0.09 | |
| Interpretation Method × Data Type | 0.67 | (2, 88) | 0.516 | 0.02 |
Technology acceptance variables
As shown in Table 7, analysis revealed a significant main effect of the interpretation method on performance expectancy, F(2, 88) = 13.92, p < 0.001, η2 = 0.24. LLM interaction was rated highest (M = 5.8, SD = 0.9), followed by in-person consultation (M = 5.5, SD = 1.0) and phone consultation (M = 4.9, SD = 1.1). A significant interaction with data type, F(2, 88) = 8.45, p < 0.01, η2 = 0.16, indicated that LLM was perceived as most useful for blood work but less so for imaging interpretation. For effort expectancy, a significant main effect of the interpretation method was observed, F(2, 88) = 19.37, p < 0.001, η2 = 0.31. LLM interaction was rated highest (M = 6.1, SD = 0.8), followed by in-person consultation (M = 5.3, SD = 1.0) and phone consultation (M = 4.8, SD = 1.1). All pairwise comparisons were significant (p < 0.05). For the behavioral intention to use, results showed a significant main effect of the interpretation method, F(2, 88) = 12.83, p < 0.001, η2 = 0.23. Participants reported the highest intention to use LLM interaction (M = 5.7, SD = 1.0), followed by in-person consultation (M = 5.4, SD = 1.1) and phone consultation (M = 4.5, SD = 1.3). A significant interaction with the data type, F(2, 88) = 9.12, p < 0.01, η2 = 0.17, indicated a greater intention to use an LLM for blood work than for imaging interpretation.
Table 7.
ANOVA results for technology acceptance variables
| Variable | Effect | F | df | p | η2 |
|---|---|---|---|---|---|
| Performance Expectancy | Interpretation Method | 13.92 | (2, 88) | <0.001 | 0.24 |
| Data Type | 5.36 | (1, 44) | 0.025 | 0.11 | |
| Interpretation Method × Data Type | 8.45 | (2, 88) | <0.01 | 0.16 | |
| Effort Expectancy | Interpretation Method | 19.37 | (2, 88) | <0.001 | 0.31 |
| Data Type | 2.94 | (1, 44) | 0.093 | 0.06 | |
| Interpretation Method × Data Type | 0.51 | (2, 88) | 0.603 | 0.01 | |
| Behavioral Intention to Use | Interpretation Method | 12.83 | (2, 88) | <0.001 | 0.23 |
| Data Type | 4.16 | (1, 44) | 0.047 | 0.09 | |
| Interpretation Method × Data Type | 9.12 | (2, 88) | <0.01 | 0.17 |
Additional outcome variables
As shown in Table 8, a significant main effect of the interpretation method on anxiety was found, F(2, 88) = 15.63, p < 0.001, η2 = 0.26. In-person consultation resulted in the lowest anxiety (M = 2.3, SD = 0.8), followed by LLM interaction (M = 2.9, SD = 1.0) and phone consultation (M = 3.6, SD = 1.1). All pairwise comparisons were significant (p < 0.05). For perceived accuracy, analysis revealed a significant main effect of the interpretation method, F(2, 88) = 17.25, p < 0.001, η2 = 0.28. In-person consultation was rated highest (M = 6.2, SD = 0.7), followed by phone consultation (M = 5.5, SD = 0.9) and LLM interaction (M = 5.2, SD = 1.0). For imaging data, this difference was more pronounced, F(2, 88) = 8.94, p < 0.01, η2 = 0.17. For trust in interpretation, a significant main effect of the interpretation method was observed, F(2, 88) = 20.18, p < 0.001, η2 = 0.31. In-person consultation generated the highest trust (M = 6.0, SD = 0.8), followed by phone consultation (M = 5.3, SD = 1.0) and LLM interaction (M = 4.6, SD = 1.2). All pairwise comparisons were significant (p < 0.01). For time efficiency, results showed a significant main effect of the interpretation method, F(2, 88) = 42.75, p < 0.001, η2 = 0.49. LLM interaction was the most time-efficient (M = 12.3 minutes, SD = 4.2), followed by phone consultation (M = 18.7 minutes, SD = 6.1) and in-person consultation (M = 24.5 minutes, SD = 7.3). All pairwise comparisons were significant (p < 0.001).
Table 8.
ANOVA results for additional outcome variables
| Variable | Effect | F | df | p | η2 |
|---|---|---|---|---|---|
| Anxiety | Interpretation Method | 15.63 | (2, 88) | <0.001 | 0.26 |
| Data Type | 3.72 | (1, 44) | 0.060 | 0.08 | |
| Interpretation Method × Data Type | 0.49 | (2, 88) | 0.614 | 0.01 | |
| Perceived Accuracy | Interpretation Method | 17.25 | (2, 88) | <0.001 | 0.28 |
| Data Type | 5.31 | (1, 44) | 0.026 | 0.11 | |
| Interpretation Method × Data Type | 8.94 | (2, 88) | <0.01 | 0.17 | |
| Trust in Interpretation | Interpretation Method | 20.18 | (2, 88) | <0.001 | 0.31 |
| Data Type | 3.95 | (1, 44) | 0.053 | 0.08 | |
| Interpretation Method × Data Type | 1.47 | (2, 88) | 0.235 | 0.03 | |
| Time Efficiency | Interpretation Method | 42.75 | (2, 88) | <0.001 | 0.49 |
| Data Type | 2.83 | (1, 44) | 0.099 | 0.06 | |
| Interpretation Method × Data Type | 0.78 | (2, 88) | 0.462 | 0.02 |
Post-hoc comparisons
Table 9 presents the results of post hoc comparisons for selected key variables, showing the specific differences between interpretation methods with the Bonferroni correction for multiple comparisons. These pairwise comparisons reveal several important patterns in how the interpretation methods differ. For diagnostic comprehension, both LLM interaction and in-person consultation significantly outperformed phone consultation (p < 0.001), but did not differ significantly from each other, suggesting that LLMs can match the effectiveness of in-person explanations for knowledge acquisition. For measures of patient empowerment, LLM interaction demonstrated superior performance in perceived control (mean difference = 0.7, p < 0.01 compared to in-person; mean difference = 1.4, p < 0.001 compared to phone), which aligns with the self-directed nature of AI-mediated interactions. However, in-person consultation maintained significant advantages in decision confidence and trust, with the largest difference observed between in-person and LLM approaches for trust in interpretation (mean difference = 1.4, p < 0.001). Perhaps most striking were the time-efficiency differences, with LLM interactions requiring substantially less time than both in-person (mean difference = −12.2 minutes, p < 0.001) and phone consultations (mean difference = −6.4 minutes, p < 0.001). These comparisons highlight the complementary strengths of different interpretation methods: LLMs excel in efficiency and control, while healthcare professionals provide superior relational value through enhanced trust and confidence.
Table 9.
Post-hoc comparisons for interpretation methods (bonferroni-corrected)
| Variable | Comparison | Mean Difference | p |
|---|---|---|---|
| Diagnostic Comprehension | LLM vs. Phone | 1.3 | <0.001 |
| LLM vs. In-Person | 0.2 | 0.420 | |
| In-Person vs. Phone | 1.1 | <0.001 | |
| Information Satisfaction | In-Person vs. Phone | 1.0 | <0.001 |
| In-Person vs. LLM | 0.5 | <0.05 | |
| LLM vs. Phone | 0.5 | <0.05 | |
| Perceived Control | LLM vs. Phone | 1.4 | <0.001 |
| LLM vs. In-Person | 0.7 | <0.01 | |
| In-Person vs. Phone | 0.7 | <0.01 | |
| Decision Confidence | In-Person vs. Phone | 1.2 | <0.001 |
| In-Person vs. LLM | 0.8 | <0.01 | |
| LLM vs. Phone | 0.4 | <0.05 | |
| Trust in Interpretation | In-Person vs. LLM | 1.4 | <0.001 |
| In-Person vs. Phone | 0.7 | <0.01 | |
| Phone vs. LLM | 0.7 | <0.01 | |
| Time Efficiency | LLM vs. In-Person | −12.2 | <0.001 |
| LLM vs. Phone | −6.4 | <0.001 | |
| Phone vs. In-Person | −5.8 | <0.001 |
Note: Only selected significant post-hoc comparisons are shown for brevity
Moderation analyses
The moderation analyses in Table 10 revealed that demographic factors influence the effectiveness of different interpretation methods. Age emerged as a significant moderator of the relationship between interpretation method and information-processing outcomes (β = −0.28, p < 0.01), with the negative coefficient indicating a stronger positive effect of LLM interaction among younger participants. Further analysis of this interaction showed that while participants under 40 years demonstrated substantial improvements in diagnostic comprehension when using LLM interaction compared to phone consultation (mean difference = 1.8, p < 0.001), participants over 60 showed a considerably smaller, though still significant, improvement (mean difference = 0.7, p < 0.05). This age-related moderation may reflect differential digital literacy, cognitive processing styles, or technology familiarity across age cohorts. Notably, our follow-up analyses indicated that age-related differences were most pronounced during initial LLM interactions and diminished somewhat across repeated exposure, suggesting potential adaptation effects.
Table 10.
Moderation effects of demographic variables
| Moderator | Relationship | Β | p |
|---|---|---|---|
| Age | Interpretation Method → Information Processing | −0.28 | <0.01 |
| Age | Interpretation Method → Behavioral Intention | −0.23 | <0.05 |
| Education Level | Interpretation Method → Patient Empowerment | 0.32 | <0.01 |
| Education Level | Interpretation Method → Diagnostic Comprehension | 0.29 | <0.01 |
| Prior Health Literacy | Interpretation Method → Trust | 0.25 | <0.05 |
| Prior Health Literacy | Interpretation Method → Anxiety | −0.21 | <0.05 |
| Gender | Interpretation Method → Information Processing | 0.15 | 0.186 |
| Gender | Interpretation Method → Patient Empowerment | 0.11 | 0.243 |
Note: Negative β values indicate stronger effects for younger participants or those with lower health literacy
Education level demonstrated an even more pronounced moderating effect on the relationship between interpretation method and patient empowerment variables (β = 0.32, p < 0.01). Participants with graduate-level education reported significantly greater increases in perceived control when using LLM interaction compared to those with high school education (interaction contrast = 0.87, p < 0.01). This educational disparity manifested most strongly in self-efficacy measures, where college-educated participants demonstrated a 26% greater improvement in health self-efficacy scores with LLM interaction relative to those without college education. Additional analyses revealed that these education-based differences persisted even after controlling for prior technology experience and health literacy, suggesting that formal educational attainment may influence information-processing strategies and self-directed learning capabilities independently of domain-specific knowledge. Previous research has similarly identified education as a critical moderator of patient response to digital health interventions, with higher education associated with greater activation through technology-mediated information [97, 98].
Health literacy showed a significant but more modest moderating effect on the relationship between interpretation method and trust outcomes (β = 0.25, p < 0.05), with participants possessing higher baseline health literacy demonstrating greater trust in LLM-provided interpretations. Interestingly, while gender appeared in our initial analyses to moderate the relationship between interpretation method and information processing (β = 0.15, p = 0.186), this effect did not reach statistical significance after controlling for age and education, suggesting that these demographic factors may be more fundamental determinants of the effectiveness of interpretation methods. These moderation findings collectively highlight the importance of considering individual differences when designing and implementing LLM-augmented medical interpretation services, as the benefits are unequally distributed across demographic segments, potentially with concerning implications for health equity.
Structural equation modeling results
Table 11 shows the model fit indices. The model showed a good fit to the data: χ2 (124) = 198.36, p < 0.001; RMSEA = 0.058 (90% CI: 0.043, 0.074); CFI = 0.944; TLI = 0.932; SRMR = 0.052. Collectively, these fit statistics demonstrate that our integrated theoretical framework—linking interpretation methods to information processing, patient empowerment, and technology acceptance variables—validly represents the relationships in the data. The model successfully captures the complex causal pathways through which different interpretation methods influence patient experiences and behavioral intentions with minimal unexplained variance.
Table 11.
Structural equation model fit indices
| Model Fit Indices | Value |
|---|---|
| χ2 (df) | 198.36 (124) |
| p-value | <0.001 |
| RMSEA (90% CI) | 0.058 (0.043, 0.074) |
| CFI | 0.944 |
| TLI | 0.932 |
| SRMR | 0.052 |
The integrated theoretical model was tested using structural equation modeling, and the results are presented in Table 12.
Table 12.
Path coefficients
| Path | Standardized Coefficient (β) | p |
|---|---|---|
| Interpretation Method → Information Processing | 0.39 | <0.001 |
| Information Processing → Patient Empowerment | 0.47 | <0.001 |
| Patient Empowerment → Technology Acceptance | 0.52 | <0.001 |
| Information Processing → Technology Acceptance | 0.31 | <0.01 |
| Interpretation Method → Technology Acceptance (Direct) | 0.24 | <0.01 |
| Interpretation Method → Technology Acceptance (Indirect) | 0.35 | <0.001 |
| Interpretation Method → Technology Acceptance (Total) | 0.59 | <0.001 |
The structural equation modeling analysis yielded several significant pathways, providing empirical support for our integrated theoretical framework. The interpretation method emerged as a significant predictor of information-processing variables (β = 0.39, p < 0.001), indicating that the mode of interpreting medical data substantially influences how effectively patients process and comprehend the information presented. These information-processing variables, in turn, demonstrated a strong positive relationship with patient empowerment variables (β = 0.47, p < 0.001), suggesting that enhanced comprehension and reduced cognitive load contribute significantly to patients’ sense of control and decision-making confidence.
The model further revealed that patient empowerment variables were strong predictors of technology acceptance outcomes (β = 0.52, p < 0.001), confirming the theoretical proposition that psychological empowerment facilitates positive attitudes toward technology adoption. In particular, information-processing variables also directly affected technology acceptance (β = 0.31, p < 0.01), indicating that cognitive factors influence adoption intentions through multiple pathways. Perhaps most significantly, the analysis confirmed a substantial indirect effect of the interpretation method on technology acceptance, mediated sequentially by information processing and patient empowerment (β = 0.35, p < 0.001). This mediation pathway, combined with the direct effect, produced a robust total effect (β = 0.59, p < 0.001), establishing that the method of interpretation substantially influences patients’ intention to adopt and use interpretation technologies through both cognitive and psychological mechanisms.
Qualitative results
The qualitative component of this study employed the grounded theory methodology outlined by Corbin and Strauss [99], utilizing a systematic approach with open, axial, and selective coding techniques to analyze data from semi-structured interviews. Following each interpretation scenario, participants engaged in brief structured interviews, and a purposively selected subset of 15 participants participated in in-depth follow-up interviews lasting 45–60 minutes. All interviews were audio-recorded, transcribed verbatim, and analyzed using NVivo 14 software to facilitate data management and coding. The analytical process began with open coding, in which two independent researchers examined the interview transcripts line by line to identify key concepts, properties, and dimensions related to participants’ experiences across different interpretation modalities. This initial coding phase generated 87 distinct codes that captured various aspects of participants’ experiences. The research team then proceeded to axial coding, establishing relationships among the initial codes by identifying causal conditions, contexts, and consequences for each phenomenon. This process involved a constant comparison between emerging categories and the theoretical frameworks of UTAUT, Patient Empowerment Theory, and Information Processing Theory to identify conceptual linkages [100].
The final phase, selective coding, involved integrating and refining the categories to develop higher-level theoretical constructs that coherently explained participants’ experiences across interpretation scenarios. This iterative process continued until theoretical saturation was achieved, with no new properties or dimensions emerging from additional data analysis [101]. Several validation strategies were employed to ensure analytic rigor, including investigator triangulation, member checking, and negative case analysis to actively seek contradictory evidence [102, 103]. Table 13 presents the five key themes identified through our grounded theory analysis, along with descriptions of each theme emerging from participants’ experiences with different medical data interpretation modalities.
Table 13.
Key themes from qualitative analysis
| Theme | Description |
|---|---|
| Convenience and Accessibility | Participants valued the 24/7 availability and immediate responsiveness of LLM interaction, allowing them to access medical information without scheduling appointments or experiencing delays. This theme connects to reduced barriers to information access and the effort expectancy construct in UTAUT. |
| Need for Human Connection | Despite technological benefits, participants emphasized the importance of human presence, particularly for emotionally significant findings. Nonverbal cues, empathy, and reassurance from healthcare professionals were especially valued when receiving concerning or complex medical information. |
| Control over Information Pace | LLM interaction enabled participants to direct their own learning experience by controlling conversation flow, asking repeated questions without embarrassment, and processing information at their preferred pace. This self-directed information gathering directly supported patient empowerment and agency. |
| Trust Development | Participants described a nuanced, evolving relationship with LLM interpretation, typically beginning with skepticism that gradually gave way to conditional trust based on experience. Trust remained context-dependent, with greater trust for straightforward information than for complex interpretations or recommendations. |
| Complementary Approaches | Rather than viewing interpretation methods as competing alternatives, participants consistently preferred hybrid approaches that integrate LLM accessibility with healthcare professional oversight, suggesting a model in which AI serves as a preparatory and supplemental resource to human expertise rather than a replacement. |
The open coding process revealed numerous participant statements on timing and accessibility, which were eventually categorized under “Convenience and Accessibility.” For example, one participant (P07, 34F) remarked, “Having immediate answers at any time was incredibly valuable. I could ask questions at midnight without feeling like I was bothering anyone.” These statements were linked to effort expectancy constructs within UTAUT through axial coding, establishing a theoretical relationship between accessibility and technology adoption intention. During axial coding, a significant pattern emerged linking emotional responses to the interpretation modality, particularly regarding findings. As one participant (P12, 42F) explained, “When the AI told me about an abnormal result, I understood the information but missed having a doctor read my reaction and reassure me.” This insight contributed to the “Need for Human Connection” theme, with selective coding revealing that this need was especially pronounced in medical imaging scenarios, where interpreting visual information created greater uncertainty.
The “Control over Information Pace” theme emerged from open codes related to information flow management and question-asking behaviors. A participant (P31, 29 M) stated, “I could ask the same question three different ways until I really understood it, which I’d feel embarrassed doing with a doctor.” Axial coding connected these experiences to Patient Empowerment Theory’s emphasis on agency and self-determination in healthcare interactions. Through selective coding, “Trust Development” emerged as a central phenomenon with complex temporal and contextual dimensions. One participant (P05, 61F) explained, “At first I didn’t trust it at all, but after seeing it correctly explain several lab values, I started to trust it for straightforward information, though I still wanted a doctor for the bigger picture.” This progressive development of trust provides context for the lower trust scores in our quantitative analysis and suggests that trust may increase with familiarity.
The most theoretically significant theme emerging from selective coding was “Complementary Approaches,” representing participants’ synthesis of their experiences across all modalities. One participant (P19, 47 M) suggested, “The ideal would be using the AI first to understand the basics and prepare questions, then meeting with my doctor for the important discussions and decisions.” This integrated perspective suggests that implementation models positioning LLMs as complementary rather than replacement technologies may achieve optimal outcomes. These qualitative findings, developed through the rigorous application of grounded theory, provide contextual depth to our quantitative results and highlight the complex, multifaceted nature of patients’ experiences across different interpretation modalities. The theoretical constructs identified through this analysis offer valuable insights for healthcare implementation strategies that leverage the unique strengths of both human and AI-mediated medical data interpretation.
Discussion
Key findings
This study investigated the comparative effects of healthcare professional-led and LLM-augmented medical data interpretation across blood work and imaging modalities. The results provide a clearer picture of strengths and limitations across interpretation methods, with significant implications for the future integration of LLMs into healthcare communication. Our mixed-methods analysis, combining quantitative performance metrics and qualitative experience data, provides a comprehensive understanding of how different interpretation modalities influence patient outcomes.
The findings demonstrate that LLM interaction offers distinct advantages in enhancing diagnostic comprehension (mean difference = 1.3 compared to phone consultation, p < 0.001), reducing cognitive load (mean difference = −1.3 compared to phone consultation, p < 0.001), increasing perceived control (mean difference = 1.4 compared to phone consultation, p < 0.001), and improving time efficiency (mean difference = −12.2 minutes compared to in-person consultation, p < 0.001). These benefits were particularly pronounced for blood work interpretation, where structured data presentation and reference ranges facilitated effective AI-mediated explanation. Our results align with studies’ emphasis on the value of self-paced, interactive learning experiences that optimize cognitive load [59, 104] and the patient empowerment theory’s focus on increased control and self-efficacy in health information management [105]. However, healthcare professional-led interpretation, particularly in-person consultation, maintained significant advantages in fostering trust (mean difference = 1.4 compared to LLM, p < 0.001), reducing anxiety (mean difference = −0.6 compared to LLM, p < 0.05), and enhancing perceived accuracy (mean difference = 1.0 compared to LLM, p < 0.001) and decision confidence (mean difference = 0.8 compared to LLM, p < 0.01). These findings corroborate previous research highlighting the importance of human connection, nonverbal communication, and empathetic presence in healthcare communication [106]. The observed difference in performance across data types, with a more pronounced preference for human interpretation in medical imaging scenarios, suggests that data complexity and visualization requirements moderate the effectiveness of different interpretation methods, a finding consistent with emerging research on AI capabilities across medical data modalities [107, 108].
Our structural equation modeling results further validate the integrated theoretical framework, confirming significant pathways from the interpretation method through information processing (β = 0.39, p < 0.001) and patient empowerment (β = 0.47, p < 0.001) to technology acceptance outcomes (β = 0.52, p < 0.001). These pathways demonstrate how interpretation modality influences cognitive processes, subsequently impacting psychological empowerment and ultimately determining technology acceptance. Identifying age, education, and health literacy as significant moderators provides important context for understanding how demographic factors influence the effectiveness of interpretation methods, with younger and more educated participants showing stronger benefits from LLM-augmented interpretation [9, 109].
Integrating our findings through the lens of our theoretical framework, we propose a comprehensive explanatory model that distinguishes between instrumental and relational values as provided by different interpretation methods. LLM interaction demonstrates particular efficacy in delivering instrumental value—enhancing comprehension, control, and efficiency—aligning with cognitive processing and the functional aspects of empowerment. In contrast, healthcare professional consultation, particularly in-person interaction, provides relational value, enhancing trust, confidence, and emotional support, thereby addressing affective aspects of information processing and the psychological dimensions of empowerment.
Comparison with previous research
Our findings both confirm and extend previous research on LLM applications in healthcare. The high diagnostic comprehension achieved through LLM interaction aligns with recent findings that LLMs can effectively communicate complex medical information to laypeople [110]. Another study found that ChatGPT generated largely accurate information for diverse medical queries, as judged by academic physicians [111]. However, our observation of differences between blood work interpretation (where LLMs excelled) and imaging interpretation (where healthcare professionals maintained advantages) extends this work by suggesting important boundary conditions based on data complexity and modality, a distinction not previously emphasized in comparative studies.
The concept of user trust in LLMs has been a pressing point in recent studies [112]. The pattern of trust results, with participants trusting healthcare professionals more than LLMs despite higher satisfaction and performance expectancy of LLMs, is consistent with the algorithm aversion phenomenon documented by Longoni, Bonezzi [113] in healthcare contexts. Their experimental studies demonstrated that patients preferred human providers even when algorithm performance was demonstrably superior, a finding they attributed to patients’ belief that human providers are uniquely capable of understanding their subjective experiences. Our results extend this work by demonstrating that this aversion persists specifically for trust-related variables while not affecting variables related to comprehension and perceived control. This supports Lankton, McKnight [114] conceptual distinction between cognitive and affective trust dimensions in human-AI interactions.
Contrary to some previous research suggesting that healthcare AI might increase patient anxiety [115–117], our findings indicate that LLM interaction resulted in lower anxiety (M = 2.9, SD = 1.0) than phone consultation (M = 3.6, SD = 1.1, p < 0.05), though higher than in-person consultation (M = 2.3, SD = 0.8, p < 0.05). This discrepancy may reflect the natural language capabilities of contemporary LLMs that create more conversational, contextually sensitive interactions compared to earlier rule-based healthcare AI tools. Recent work similarly found that conversational LLMs elicited less anxiety than traditional clinical decision support systems, attributing this difference to enhanced communication fluidity and personalization [118]. This instrumental-relational value distinction builds upon and extends recent theoretical work proposing a dual-pathway model of healthcare information processing, suggesting that patients engage with medical information through both epistemic (knowledge-focused) and affective (emotion-focused) cognitive channels [119, 120]. Our findings provide empirical validation for this conceptualization, demonstrating that LLM interaction optimizes epistemic processing (as evidenced by higher comprehension scores) while healthcare professional consultation facilitates affective processing (as demonstrated by enhanced trust and reduced anxiety). Similarly, this distinction aligns with the functional-experiential dichotomy in medical communication [37, 121], differentiating between information transactions and relational experiences in clinical interactions.
Our qualitative finding that participants preferred hybrid approaches aligns with the concept of “augmented intelligence” rather than artificial intelligence in healthcare, where technology enhances rather than replaces human capabilities [122]. This conceptualization has gained significant traction in recent literature, as recent studies documented successful implementations of human-AI collaborative systems across multiple healthcare contexts [123]. Our findings provide empirical support for this approach, with participants explicitly articulating value in using LLMs for initial learning and preparation, followed by human consultation for complex decision-making and emotional support. These findings have significant implications for implementation models in healthcare settings. Rather than positioning LLMs and healthcare professionals as competing alternatives, our results suggest that optimal patient outcomes may be achieved through complementary integration that leverages the unique strengths of each modality. This aligns with recent implementation research indicating successful hybrid models where AI tools augmented rather than replaced clinician communication across multiple healthcare contexts [124]. Also, the qualitative finding that participants envisioned using LLMs for initial learning before healthcare professional consultations suggests a sequential integration model that maximizes both instrumental and relational value. Such an approach would allow patients to benefit from the on-demand, self-paced learning facilitated by LLMs while preserving the trust, emotional support, and personalized guidance healthcare professionals provide. This implementation model aligns with the concept of preparatory AI engagement, which has demonstrated improved consultation efficacy through pre-visit patient education and question preparation [125].
Recent studies have also begun to explore the differential effectiveness of AI tools across medical data types. For example, the evidence suggests that structured data tasks generally achieve higher, more consistent accuracy (often 80–90% or higher) due to standardized inputs and a lower reliance on spatial reasoning [126]. Image-based diagnostics, despite advancements in deep learning (e.g., CNNs, Vision Transformers), face challenges such as variability in imaging quality, the need for explainable reasoning, and difficulties with rare or context-dependent conditions, which can reduce reliability in complex cases [127]. Similarly, Tu, Palepu [128] recognized stronger patient acceptance of AI consultation for laboratory result interpretation than for radiological image analysis. Our findings extend this emerging pattern by providing a controlled comparative analysis across both data types and communication modalities, offering a more comprehensive understanding of the contextual factors influencing AI effectiveness in healthcare communication. The observed interaction between the interpretation method and data type further validates this conceptual model. Blood work interpretation, being fundamentally quantitative and reference-range-based, primarily requires instrumental value (a clear explanation of numerical values), and LLMs demonstrated superior performance. In contrast, imaging interpretation involves greater visual complexity, interpretive uncertainty, and emotional significance, increasing the importance of relational value provided by healthcare professionals. This pattern aligns with recent empirical work showing differential patient preferences for AI versus human interpretation, depending on the complexity of medical data and its emotional valence [4, 129].
Our study’s moderation effects of age and education align with digital health equity research, which identified significant disparities in AI-mediated healthcare benefit distribution across demographic groups [130]. Some studies found that older adults and those with lower educational attainment consistently derived less benefit from digital health interventions [131, 132], a pattern our moderation analysis confirms specifically for LLM-mediated medical interpretation. These findings highlight the importance of considering diverse user capabilities when implementing AI-augmented healthcare communication systems [40].
Theoretical implications
The integrated theoretical model proposed in this study receives substantial empirical support and makes significant contributions to multiple theoretical domains in healthcare information systems and patient-centered care. The structural equation modeling results confirm a hierarchical progression from the interpretation method through information processing and patient empowerment to technology acceptance, thereby validating the underlying mechanisms of our conceptual framework. This empirical validation extends the UTAUT in several meaningful ways. First, our findings go beyond UTAUT’s traditional direct-effect model by showing how cognitive and psychological processes activated by different technologies mediate their eventual acceptance [133]. A study by Maruping, Bala [134] called for this extension, arguing that understanding the black box of cognitive mechanisms between technology characteristics and acceptance outcomes represents a critical frontier in technology adoption research. Our study addresses this call by empirically mapping these intervening processes and demonstrating that information-processing efficacy is a necessary precursor to both psychological empowerment and subsequent technology acceptance.
Second, our findings contribute to the emerging theoretical discourse on contextual technology acceptance by documenting how data type moderates the effectiveness of different interpretation methods. This aligns with the conceptualization of contextual technology acceptance [135], which proposes that adoption factors vary systematically across use contexts and information types [136]. Our demonstration of differential effectiveness between blood work and imaging interpretation provides empirical validation of this theoretical perspective, showing that technology acceptance is not uniform but context-dependent on the complexity and emotional significance of the information being processed. The study also advances patient empowerment theory in several important dimensions. While traditional conceptualizations of patient empowerment have emphasized provider-facilitated activation [12, 51], our findings demonstrate that technological tools can independently enhance perceived control and self-efficacy when they enable self-directed information exploration [137]. Specifically, the finding that LLM interaction increased perceived control even beyond in-person consultation challenges the longstanding assumption that human providers are inherently more empowering and suggests new pathways for patient activation. This finding aligns with recent work suggesting that technology-mediated autonomy represents a distinct pathway to empowerment, separate from traditional provider-mediated approaches [138, 139]. Similarly, a study identified informational self-determination as a distinct form of patient empowerment enabled by digital health tools that allow patients to control the pace and depth of their engagement with health information [140]. Our empirical validation of these theoretical propositions advances our understanding of how AI-mediated interactions can create unique forms of patient empowerment not achievable through traditional clinical encounters.
In the context of information processing theory, our results offer several theoretical extensions. First, the finding that LLM interaction reduced cognitive load compared to both phone and in-person consultations validates and extends Sweller [55]‘s cognitive load theory of educational technology by demonstrating how interactive, adaptive information presentation can optimize cognitive resource allocation. Recent theoretical work on conversational information processing predicted this effect, proposing that dialogue-based information delivery allows for dynamic cognitive load management through iterative question-answer cycles [141, 142], a prediction our findings empirically confirm. Second, our results contribute to understanding the relationship between information modality and comprehension. The LLM’s ability to tailor explanations to user questions and provide contextual information reduces cognitive load while enhancing comprehension, especially for structured data such as blood work results. This finding extends multimedia learning theory by suggesting that interactivity may be as important as multimodality in optimizing information processing [143, 144]. Recent studies on adaptive conversational agents proposed that educational systems should dynamically adjust explanation depth based on learner feedback, a critical mechanism for enhancing comprehension of complex information [145, 146]. Our empirical demonstration of LLMs’ effectiveness validates this theoretical proposition.
Finally, our integrated theoretical framework offers a novel synthesis of previously disparate theoretical traditions. By demonstrating empirical links among information processing, patient empowerment, and technology acceptance, we contribute to the development of more comprehensive models of patient-technology interaction in healthcare settings. This integration aligns with the research avenue on theoretical coherence in digital health [147, 148] by empirically validating interconnections between cognitive, psychological, and behavioral dimensions of health technology experiences.
Practical implications
Our findings yield substantial implications for healthcare organizations implementing LLM-augmented medical data interpretation services. The differential effectiveness observed across data types suggests that a context-sensitive approach to LLM integration is more reasonable. Specifically, LLMs demonstrate particular efficacy in interpreting structured, numerical medical data, such as blood work results, while more complex visual data, such as medical imaging, may continue to benefit from healthcare professionals’ expertise. This data-type dependency indicates that healthcare systems should allocate interpretation resources strategically based on content complexity rather than implement one-size-fits-all solutions. Furthermore, our research strongly suggests that conceptualizing LLMs as complementary tools rather than replacements for healthcare professionals offers the most promising path to implementation, emphasizing technology-human collaboration rather than substitution [149]. Models in which LLMs provide initial or supplementary interpretation under professional oversight align with both our quantitative results and participants’ expressed preferences in qualitative interviews. This complementary approach has demonstrated effectiveness in pilot implementations of AI across various healthcare settings [150]. It offers a framework for healthcare organizations to preserve the relational aspects of care while enhancing informational efficiency.
The significant moderation effects of demographic variables on interpretation outcomes demand careful consideration in implementation planning. Our observation that age and education substantially moderate the relationship between interpretation method and outcome variables indicates that healthcare systems must develop targeted support strategies for older and less-educated patients to ensure equitable distribution of benefits [151]. These findings align with broader digital health equity concerns and suggest that LLM interface design and deployment should incorporate accessibility features and additional support mechanisms for vulnerable populations. Addressing the trust gap observed in our results represents another critical implementation consideration. Healthcare organizations should incorporate trust-building elements into LLM interfaces, including transparent sourcing of medical information, clear confidence indicators for the interpretations provided, and seamless integration pathways to healthcare professional services when needed. Institutional endorsement of AI technologies has significantly enhanced user trust [152], suggesting that healthcare systems should explicitly communicate their validation processes for deployed LLM systems. Such features may help mitigate the trust differential currently favoring human providers while leveraging the cognitive and efficiency benefits of LLM interaction.
Finally, the substantial time-efficiency advantage of LLM interaction (nearly a 50% reduction compared to in-person consultation) offers healthcare systems an opportunity to optimize resource allocation while enhancing patient access to medical information. In practice, this efficiency gain could allow healthcare professionals to focus on complex cases and emotionally significant findings where their unique human capabilities provide maximal value, while routine interpretation tasks might be augmented or partially automated through LLM integration [40]. Implementation strategies should incorporate robust evaluation frameworks to monitor both the intended and unintended consequences of LLM integration. A multidimensional evaluation approach encompassing technical performance, workflow integration, patient outcomes, and health equity dimensions can serve as a framework well aligned with our study’s findings on differential effectiveness across patient populations and data types. Additionally, governance structures should be established to oversee continuous monitoring and refinement of LLM-augmented interpretation services. Collectively, these implications suggest that thoughtfully designed and strategically implemented LLM-augmented interpretation services have the potential to enhance healthcare system efficiency, improve information accessibility, and promote patient empowerment without sacrificing the essential human elements of healthcare communication. The implementation pathway should be guided by evidence-based principles of user-centered design, equity-focused deployment, strategic task allocation, and continuous performance evaluation to realize these potential benefits. To illustrate how these complementary strengths might be integrated in practice, Fig. 3 presents a conceptual model of hybrid implementation:
Fig. 3.
Integration of instrumental and relational strengths in hybrid interpretation models
Hybrid interpretation models are based on the principle that LLMs and professional consultation represent complementary approaches with distinct strengths, rather than competing alternatives. Sequential implementation leverages LLM strengths in the preparation phase (comprehension, efficiency, autonomy), followed by professional consultation that provides the clinical judgment, trust, and personalization that patients value. This integrated approach addresses both instrumental patient needs (understanding medical information) and relational patient needs (confidence, trust, professional guidance).
Figure 3 illustrates how the strengths of LLMs and professional consultation can be strategically integrated into hybrid interpretation models. LLMs offer instrumental strengths: high efficiency, enhanced comprehension through plain language, perceived control through self-directed learning, and accessibility through unlimited, non-judgmental repetition. Healthcare professionals provide complementary relational strengths: clinical authority grounded in expertise; trust and confidence through professional credibility; shared decision-making capabilities; and holistic care responsive to individual patient needs. Rather than replacing one approach with another, hybrid models leverage both sets of strengths simultaneously. Sequential implementation (LLM preparation followed by professional consultation) optimizes patient comprehension while maintaining professional oversight. The result is a synergistic approach that maximizes both efficiency and quality outcomes. For detailed implementation strategies across different clinical contexts, see Appendix C.
Limitations and future research
Limitations
Several limitations need to be considered when interpreting the results of this study. First, while the experimental approach offers strong internal validity, the hypothetical nature of the medical data may not fully replicate the emotional significance of receiving actual personal health information. We used standardized blood work results and radiological images that did not belong to the participants, which allowed for experimental control but potentially diminished the personal relevance and emotional investment typically associated with interpreting one’s own medical data. Second, our inclusion criteria (adults aged 20–70 years with English proficiency, basic digital literacy, prior experience with medical test results, and a specific regional location) and recruitment methods yielded a sample of 45 participants with specific demographic characteristics. While this sample was strategically diverse across key variables, it may not be fully representative of the broader population, particularly those with lower educational attainment or from non-urban settings. Additionally, the requirement for basic digital literacy may obscure barriers for less technologically adept individuals, who might face different challenges with LLM-mediated interpretation. Third, our use of ChatGPT-4o as the LLM representative in this study, while reflecting current real-world patient practices, presents inherent limitations. As a general-purpose model rather than a specialized healthcare AI system, it may perform differently from specialized medical AI systems. Furthermore, the rapid advancement of AI capabilities means our findings may not generalize to future iterations of LLMs or to alternative models with different architectures and training approaches. Fourth, assessing immediate reactions to interpretation scenarios does not capture longitudinal effects such as delayed comprehension or evolving trust relationships. Fifth, while necessary for experimental control, standardized protocols for healthcare professionals’ consultations may not fully capture the complexity and variability of real-world clinical communication. Sixth, while objective measures were included (e.g., a comprehension test, time efficiency), many outcomes relied on self-reported perceptions, which may be subject to response bias, particularly given the novelty effect often associated with new technologies in healthcare settings.
Finally, we also need to acknowledge potential biases and methodological considerations. The potential biases introduced by order effects and scenario sequencing; novelty effects with LLM technology; pre-existing biases towards AI and healthcare professionals; experimenter effects in healthcare professional consultations; and ecological validity and hypothetical scenarios. Order effects and scenario sequencing: Although this study employed a Latin square design to randomize the order of scenario presentation and mitigate potential order effects, residual order biases may persist. Specifically, participants may have performed better on later scenarios due to increased familiarity with the experimental protocol, or conversely, fatigue effects may have diminished performance on later scenarios. Although randomization helps distribute these effects across conditions, systematic patterns cannot be entirely ruled out. Some participants may have demonstrated learning effects when encountering similar interpretation modalities sequentially (e.g., phone consultation followed by in-person consultation), potentially biasing comparisons between professional consultation modalities. Future research employing between-subjects designs for interpretation modality could address this limitation.
Novelty effects with LLM technology: Given that 42.2% of participants reported no prior experience with LLMs, novelty effects likely influenced responses to LLM scenarios. Participants may have experienced heightened engagement, skepticism, or curiosity regarding the LLM interaction compared to more familiar professional consultation modalities. This novelty effect could artificially inflate or deflate outcome measures (comprehension, satisfaction, adoption intentions) depending on whether the novelty was perceived positively or negatively. The finding that repeated exposure to LLMs showed some adaptation effects (differences diminished across repeated scenarios) suggests that initial novelty influenced early responses. Results may not generalize to scenarios where patients have greater familiarity with LLM interpretation tools.
Pre-existing biases towards AI and healthcare professionals: Participants likely entered the study with attitudes and biases regarding both AI and healthcare professionals that could not be fully controlled for. Some participants may have held technology-optimistic views that enhanced their perception of LLM capabilities, while others may have held technology-skeptical views that diminished perceptions. Similarly, pre-existing trust in or skepticism toward healthcare professionals influenced responses to the conditions of professional consultation. These baseline attitudes, though potentially balanced across conditions through randomization, may have introduced systematic biases that affected comparative judgments. Participants’ prior experiences with AI, healthcare systems, and medical decision-making likely influenced their interpretation of all scenarios. The study did not measure baseline AI attitudes or healthcare trust prior to scenario exposure, limiting the ability to statistically control for these potential confounds.
Experimenter effects in healthcare professional consultations: Although physicians were trained to standardize communication, subtle differences in rapport-building, nonverbal communication, and physician demeanor may have influenced participant responses differently across the two physicians in each modality (phone vs. in-person). Participants may have responded differently to individual physicians’ communication styles, independent of the interpretation method being tested. Additionally, knowledge that the study involved AI comparison may have unconsciously motivated physicians to communicate with particular emphasis or clarity, creating subtle experimenter effects.
Ecological validity and hypothetical scenarios: The use of standardized, anonymized medical data rather than participants’ own personal health information, while methodologically necessary, reduced ecological validity. Participants may have engaged less emotionally or cognitively with scenarios involving others’ medical data than with interpreting their own results. Personal health information typically carries emotional weight and stakes that may alter comprehension, trust development, and technology acceptance in ways not captured in hypothetical scenarios. Results may not generalize to real-world clinical contexts in which patients interpret personally significant medical data.
Future research directions
Building on this study’s findings and addressing its limitations, several promising directions for future research emerge. First, future studies can extend our study by conducting longitudinal investigations to examine how repeated interactions with LLM interpreters affect trust development, learning patterns, and health behavior over time. Such studies could assess whether the initial trust differential between healthcare professionals and LLMs diminishes with extended exposure and familiarity. Second, future research can utilize participants’ actual personal medical data rather than standardized cases to enhance ecological validity. Examining how patients interpret their own medical results would provide more authentic insights into the emotional and psychological dimensions of interpreting medical data that may not be fully captured by hypothetical scenarios. This approach would likely reveal additional complexities in patient responses, particularly concerning or abnormal findings with direct personal health implications. Third, experimental studies can systematically evaluate hybrid integration models that combine LLMs with healthcare professionals’ interpretation. These investigations could assess various implementation approaches, such as sequential (LLM followed by provider), parallel (both sources simultaneously), or integrated (provider-guided LLM use), to identify optimal configurations for different clinical contexts and patient populations. Such research would provide evidence-based guidance for healthcare systems implementing complementary interpretation models. Fourth, extending this comparative analysis to additional types of medical data beyond blood work and radiographic images would further define the boundaries of LLM effectiveness. Genetic testing results, electrophysiological data (e.g., ECGs and EEGs), pathology reports, and continuous monitoring data are promising domains for comparative assessment. These investigations could validate or refine our instrumental-relational value framework across diverse types of medical data. Fifth, future research can move beyond experiential measures to examine impacts on clinical outcomes. Prospective studies measuring how different interpretation methods affect treatment adherence, preventive behaviors, healthcare utilization, and health status would strengthen the evidence base for implementation decisions. This research could determine whether the cognitive and psychological benefits observed in our study translate into meaningful changes in health behavior.
Sixth, a more comprehensive investigation of how individual differences moderate the effects of interpretation methods is warranted. While our study identified age and education as significant moderators, future research should examine additional characteristics, including health literacy, technology anxiety, cultural factors, language proficiency, and disability status, to inform more personalized implementation strategies. This work is essential for ensuring equitable distribution of benefits from LLM-augmented interpretation. Seventh, developing and testing specific communication protocols for both healthcare professionals and LLMs based on our findings could optimize the effectiveness of each approach. Experimental manipulation of LLM design features, including uncertainty communication, empathetic language, visual aids, and information density, could identify specific elements that enhance patient comprehension and trust. Similarly, training interventions for healthcare professionals focused on strengthening relational communication while leveraging LLM’s instrumental benefits would support complementary implementation models. Finally, implementing and evaluating LLM interpretation in real-world clinical settings would enhance ecological validity and identify practical implementation challenges. Pragmatic trials comparing different integration approaches across diverse healthcare contexts could generate valuable insights regarding workflow integration, provider acceptance, patient utilization patterns, and clinical outcomes. This implementation of scientific research would bridge the gap between controlled experimental findings and real-world healthcare delivery, ultimately advancing evidence-based integration of LLM-augmented medical data interpretation into clinical practice.
Conclusions
This study comprehensively compares healthcare professional-led and LLM-augmented approaches to medical data interpretation, revealing distinctive patterns of strengths and limitations across different data types and outcome dimensions. By integrating perspectives on technology acceptance, patient empowerment, and information processing, we offer a novel theoretical framework for understanding the complex interplay among interpretation methods, cognitive processes, psychological outcomes, and behavioral intentions. The findings suggest that LLM-augmented interpretation offers significant advantages for enhancing patient comprehension, control, and efficiency, particularly for structured medical data such as blood test results. However, healthcare professionals’ interpretation remains an important advantage for fostering trust, confidence, and emotional support, especially for complex visual data such as medical imaging. These results point toward complementary rather than competitive roles for human and AI interpretation. As healthcare systems globally face increasing demands with constrained resources, LLMs offer promising opportunities to enhance patient access to medical information while potentially reducing clinician burden. However, their implementation should be strategic and context-sensitive, recognizing both their capabilities and limitations. The framework and evidence from this study provide healthcare leaders, technology developers, and policymakers with a foundation for designing patient-centered approaches to medical data interpretation that leverage the unique strengths of both human expertise and artificial intelligence. By advancing our understanding of how patients experience and process medical information across different interpretation modalities, this research contributes to the broader goal of creating healthcare systems that are more efficient, empowering, comprehensible, and responsive to patient needs in an increasingly digital healthcare landscape.
Electronic supplementary material
Below is the link to the electronic supplementary material.
Acknowledgements
We thank all participants for their valuable time and insights during the study sessions and interviews. We are also grateful to the panel of board-certified physicians whose expertise in general medicine and radiology ensured the clinical relevance and realism of the experimental scenarios through their careful review and validation of the study materials.
Abbreviations
- AI
Artificial intelligence
- ANOVA
Analysis of variance
- AVE
Average variance extracted
- CBC
Complete blood count
- CFA
Confirmatory factor analysis
- CFI
Comparative fit index
- CMP
Comprehensive metabolic panel
- DSES
Decision Self-Efficacy Scale
- GPT
Generative Pre-trained Transformer
- HCEI
Health Care Empowerment Inventory
- HIPAA
Health Insurance Portability and Accountability Act
- LLM
Large Language Model
- MISS-21
Medical Information Satisfaction Scale
- RMSEA
Root mean square error of approximation
- SEM
Structural equation modeling
- SPSS
Statistical Package for the Social Sciences
- SRMR
Standardized root mean square residual
- STAI
State-Trait Anxiety Inventory
- UTAUT
Unified Theory of Acceptance and Use of Technology.
Author contributions
PE conceived the idea for this study. He collected the data and conducted the data analysis. PE interpreted the findings, drafted the manuscript, and critically revised it for significant intellectual content. He also read and approved the final version of the manuscript.
Funding
This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.
Data availability
A significant portion of the data analyzed in this study is included in this published article and three Appendices. The data used and analyzed during the current study are available from the corresponding author upon reasonable request.
Declarations
Ethics approval and consent to participate
The Office of Institutional Review Board (IRB) at Florida International University (FIU) fully evaluated and approved the study. The IRB reference number is #115871. All participants provided written informed consent to take part in the research. Additionally, the study complied with the Declaration of Helsinki regarding research involving human data.
Consent for publication
Not applicable.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
- 1.Chauhan P, Bali A, Kaur S. Breaking barriers for accessible health programs: the role of telemedicine in a global healthcare transformation, in transformative approaches to patient literacy and healthcare innovation. IGI Global. 2024;283–307.
- 2.Topol EJ. High-performance medicine: the convergence of human and artificial intelligence. Nat Med. 2019;25(1):44–56. 10.1038/s41591-018-0300-7. [DOI] [PubMed] [Google Scholar]
- 3.Lian OS, Nettleton S, Grange H, Dowrick C. “I’m not the doctor; I’m just the patient”: patient agency and shared decision-making in naturally occurring primary care consultations. Patient Educ Couns. 2022;105(7):1996–2004. 10.1016/j.pec.2021.10.031. [DOI] [PubMed] [Google Scholar]
- 4.Esmaeilzadeh P, Maddah M, Mirzaei T. Using AI chatbots (e.g., CHATGPT) in seeking health-related information online: the case of a common ailment. Comput Hum Behav: Artif Humans. 2025;3:100127. 10.1016/j.chbah.2025.100127. [DOI] [Google Scholar]
- 5.He Z, Bhasuran B, Jin Q, Tian S, Hanna K, Shavor C, et al. Quality of answers of generative large language models versus peer users for interpreting laboratory test results for lay patients: evaluation study. J Med Internet Res. 2024;26:e56655. 10.2196/56655. [DOI] [PMC free article] [PubMed]
- 6.Wu Y, et al. Embracing ChatGPT for medical education: exploring its impact on doctors and medical students. Vol. 10. JMIR Medical Education; 2024. p. e52483. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Nazi ZA, Peng W. Large language models in healthcare and medical domain: a review. In: Informatics. MDPI; 2024.
- 8.AlSaad R, Abd-Alrazaq A, Boughorbel S, Ahmed A, Renault M-A, Damseh R, et al. Multimodal large language models in health care: applications, challenges, and future outlook. J Med Internet Res. 2024;26:e59505. 10.2196/59505. [DOI] [PMC free article] [PubMed]
- 9.Rajashekar NC, et al. Human-algorithmic interaction using a large language model-augmented artificial intelligence clinical decision support system. Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. 2024.
- 10.Li J, Zhou Z, Lyu H, Wang Z. Large language models-powered clinical decision support: enhancing or replacing human expertise? Intell Med. 2025;5(1):1–4. 10.1016/j.imed.2025.01.001. [DOI] [Google Scholar]
- 11.Venkatesh V, Morris MG, Davis GB, Davis FD. User acceptance of information technology: toward a unified view. Mis Q. 2003;27(3):425–78. 10.2307/30036540. [DOI] [Google Scholar]
- 12.Bravo P, Edwards A, Barr PJ, Scholl I, Elwyn G, McAllister M. Conceptualising patient empowerment: a mixed methods study. BMC Health Serv Res. 2015;15(1):1–14. 10.1186/s12913-015-0907-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Frisch A-L, Camerini L, Schulz PJ. The impact of presentation style on the retention of online health information: a randomized-controlled experiment. Health Commun. 2013;28(3):286–93. 10.1080/10410236.2012.683387. [DOI] [PubMed] [Google Scholar]
- 14.de Gomez MRC. A comprehensive introduction to healthcare data analytics. JBSHA. 2024;44–53. 10.53759/0088/JBSHA20240405. [DOI]
- 15.Tayefi M, Ngo P, Chomutare T, Dalianis H, Salvi E, Budrionis A, et al. Challenges and opportunities beyond structured data in analysis of electronic health records. WIREs Comput Stats. 2021;13(6):e 1549. 10.1002/wics.1549. [DOI]
- 16.Zhang CX, Crawford E, Marshall J, Bernard A, Walker-Smith K. Developing interprofessional collaboration between clinicians, interpreters, and translators in healthcare settings: outcomes from face-to-face training. J Interpro Care. 2021;35(4):521–31. 10.1080/13561820.2020.1786360. [DOI] [PubMed] [Google Scholar]
- 17.Hickmann E, Richter P, Schlieter H. All together now – patient engagement, patient empowerment, and associated terms in personal healthcare. BMC Health Serv Res. 2022;22(1):1116. 10.1186/s12913-022-08501-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Esmaeilzadeh P, Dharanikota S, Mirzaei T. The role of patient engagement in patient-centric health information exchange (HIE) initiatives: an empirical study in the United States. Information Technology & People; 2021. [Google Scholar]
- 19.Liu LG, Ancker JS, Creber RMM. Improving patient engagement through patient decision support. Am J Prev Med. 2021;60(3):438–41. 10.1016/j.amepre.2020.08.010. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Solomou T, Mappouras S, Kyriacou E, Constantinou I, Antoniou Z, Canciu IC, et al. Bridging language barriers in healthcare: a patient-centric mobile app for multilingual health record access and sharing. Front Digit Health. 2025;7:1542485. 10.3389/fdgth.2025.1542485. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Adeghe EP, Okolo CA, Ojeyinka OT. The role of big data in healthcare: a review of implications for patient outcomes and treatment personalization. World J Bio Pharm Health Sci. 2024;17(3):198–204. 10.30574/wjbphs.2024.17.3.0133. [DOI] [Google Scholar]
- 22.Moulaei K, Sheikhtaheri A, Fatehi F, Shanbehzadeh M, Bahaadinbeigy K. Patients’ perspectives and preferences toward telemedicine versus in-person visits: a mixed-methods study on 1226 patients. BMC Med Inf Decis Mak. 2023;23(1):261. 10.1186/s12911-023-02348-4. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Ullah MW, Rahman R, Nilima SI, Tasnim AF, Aziz MB. Health behaviors and outcomes of mobile health Apps and patient engagement in the Usa. JCC. 2024;12(10):78–93. 10.4236/jcc.2024.1210007. [DOI] [Google Scholar]
- 24.Newman B, Joseph K, Chauhan A, Seale H, Li J, Manias E, et al. Do patient engagement interventions work for all patients? A systematic review and realist synthesis of interventions to enhance patient safety. Health Expectations. 2021;24(6):1905–23. 10.1111/hex.13343. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Klein B, Perfetto EM, Oehrlein EM, Weston F, Lobban TCA, Boutin M. Measuring and demonstrating the value of patient engagement across the medicines lifecycle: a patient engagement impact measurement framework. Patient-Patient-Centered Outcomes Res. 2025;18(1):3–18. 10.1007/s40271-024-00713-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Alawiye TR. The impact of digital technology on healthcare delivery and patient outcomes. ETSN. 2024;13(2):13–22. 10.4236/etsn.2024.132002. [DOI] [Google Scholar]
- 27.Fitzpatrick PJ. Improving health literacy using the power of digital communications to achieve better health outcomes for patients and practitioners. Front Digit Health. 2023;5:1264780. 10.3389/fdgth.2023.1264780. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 28.Khalifa M, Albadawy M, Iqbal U. RETRACTED: advancing clinical decision support: the role of artificial intelligence across six domains. Comput Methods Programs Biomed Update. 2024;5:100142. 10.1016/j.cmpbup.2024.100142. [DOI] [Google Scholar]
- 29.Kunze KN, Nwachukwu BU, Cote MP, Ramkumar PN. Large language models applied to health care tasks may improve clinical efficiency, value of care rendered, research, and medical education. Arthrosc: J Arthroscopic Relat Surg. 2025;41(3):547–56. 10.1016/j.arthro.2024.12.010. [DOI] [PubMed] [Google Scholar]
- 30.Keshavarz P, Bagherieh S, Nabipoorashrafi SA, Chalian H, Rahsepar AA, Kim GHJ, et al. ChatGPT in radiology: a systematic review of performance, pitfalls, and future perspectives. Diagn Interv Imag. 2024;105(7–8):251–65. 10.1016/j.diii.2024.04.003. [DOI] [PubMed] [Google Scholar]
- 31.Jin H, Guo J, Lin Q, Wu S, Hu W, Li X. Comparative study of Claude 3.5-Sonnet and human physicians in generating discharge summaries for patients with renal insufficiency: assessment of efficiency, accuracy, and quality. Front Digit Health. 2024;6:1456911. 10.3389/fdgth.2024.1456911. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 32.Taheri A, Farhadi A, Zamanifar A. Application of GenAI in clinical administration support. In: Application of generative AI in healthcare systems. Springer; 2025. p. 91–117.
- 33.Kalidindi S, Vadana RP. Overview of the potentially transforming role of Large language models (LlMs) in creating patient-friendly summaries for radiology reports. J Radiol Clin Imag 2024. 2024;7(4):54–59. 10.26502/jrci.2809097. [DOI] [Google Scholar]
- 34.Aydin S, Karabacak M, Vlachos V, Margetis K. Large language models in patient education: a scoping review of applications in medicine. Front Med. 2024;11:1477898. 10.3389/fmed.2024.1477898. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 35.Esmaeilzadeh P. Baseline evaluation of Claude opus 4 for diabetes management: a preliminary assessment and lessons for implementation. Appl Clin Inf. 2025;16(5):1881–91. 10.1055/a-2765-6930. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 36.Singh M, et al. From text to treatment: an Overview of artificial intelligence chatbots in healthcare. 2024 11th International Conference on Computing for Sustainable Global Development (INDIACom). IEEE; 2024.
- 37.Mirzaei T, Amini L, Esmaeilzadeh P. Clinician voices on ethics of LLM integration in healthcare: a thematic analysis of ethical concerns and implications. BMC Med Inf Decis Mak. 2024;24(1):250. 10.1186/s12911-024-02656-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 38.Goodman KE, Yi PH, Morgan DJ. AI-generated clinical summaries require more than accuracy. JAMA. 2024;331(8):637–38. 10.1001/jama.2024.0555. [DOI] [PubMed] [Google Scholar]
- 39.Zhang X, Xiong Y, Miller A. ChatGPT supporting heritage language literacy in underserved communities: a neurocognitive study of sociolinguistic factors. Digit Stud Language Lit. 2025;2(1):153–77. 10.1515/dsll-2024-0011. [DOI] [Google Scholar]
- 40.Esmaeilzadeh P. Challenges and strategies for wide-scale artificial intelligence (AI) deployment in healthcare practices: a perspective for healthcare organizations. Artif Intell Med. 2024;151:102861. 10.1016/j.artmed.2024.102861. [DOI] [PubMed] [Google Scholar]
- 41.Ye C, Zweck E, Ma Z, Smith J, Katz S. Doctor versus artificial intelligence: patient and physician evaluation of Large language model responses to Rheumatology patient questions in a cross-sectional study. Arthritis Rheumatol. 2024;76(3):479–84. 10.1002/art.42737. [DOI] [PubMed] [Google Scholar]
- 42.Singhal K, et al. Toward expert-level medical question answering with large language models. Nat Med. 2025;1–8. [DOI] [PMC free article] [PubMed]
- 43.Kim S, Lee K-H, Hwang H, Yoo S. Analysis of the factors influencing healthcare professionals’ adoption of mobile electronic medical record (EMR) using the unified theory of acceptance and use of technology (UTAUT) in a tertiary hospital. BMC medical informatics and decision making. Bmc Med Inf Decis. 2015;16(1):1–12. 10.1186/s12911-016-0249-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 44.Rouidi M, Elouadi AE, Hamdoune A, Choujtani K, Chati A. TAM-UTAUT and the acceptance of remote healthcare technologies by healthcare professionals: a systematic review. Inf Med Unlocked. 2022;32:101008. 10.1016/j.imu.2022.101008. [DOI] [Google Scholar]
- 45.Hassan M, Kushniruk A, Borycki E. Barriers to and facilitators of artificial intelligence adoption in health care: scoping review. JMIR Hum Factors. 2024;11:e48633. 10.2196/48633. [DOI] [PMC free article] [PubMed]
- 46.Ren Z, et al. Healthcare copilot: eliciting the power of general llms for medical consultation. arXiv preprint arXiv: 2402.13408, 2024.
- 47.Mora MA, Sparud-Lundin C, Moons P, Bratt E-L. Definitions, instruments and correlates of patient empowerment: a descriptive review. Patient Educ Couns. 2022;105(2):346–55. 10.1016/j.pec.2021.06.014. [DOI] [PubMed] [Google Scholar]
- 48.Anderson RM, Funnell MM. Patient empowerment: myths and misconceptions. Patient Educ Couns. 2010;79(3):277–82. 10.1016/j.pec.2009.07.025. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 49.Small N, Bower P, Chew-Graham CA, Whalley D, Protheroe J. Patient empowerment in long-term conditions: development and preliminary testing of a new measure. BMC Health Serv Res. 2013;13(1):1–15. 10.1186/1472-6963-13-263. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 50.Farzandipour M, Nabovati E, Heidarzadeh Arani M, Akbari H, Sharif R, Anvari S. Enhancing asthma patients’ self-management through smartphone-based application: design, usability evaluation, and educational intervention. Appl Clin Inf. 2019;10(5):870–78. 10.1055/s-0039-1700866. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 51.Castro EM, Van Regenmortel T, Vanhaecht K, Sermeus W, Van Hecke A. Patient empowerment, patient participation and patient-centeredness in hospital care: a concept analysis based on a literature review. Patient Educ Couns. 2016;99(12):1923–39. 10.1016/j.pec.2016.07.026. [DOI] [PubMed] [Google Scholar]
- 52.Li B, et al. Beyond the waiting room: Patient’s perspectives on the conversational nuances of pre-consultation chatbots. Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. 2024.
- 53.Manalili K, Siad FM, Antonio M, Lashewicz B, Santana MJ. Codesigning person-centred quality indicators with diverse communities: a qualitative patient engagement study. Health Expectations. 2022;25(5):2188–202. 10.1111/hex.13388. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 54.Grünloh C, Myreteg G, Cajander Å, Rexhepi H. “Why do they need to check me?” patient participation through eHealth and the doctor-patient relationship: qualitative study. J Med Internet Res. 2018;20(1):e11. 10.2196/jmir.8444. [DOI] [PMC free article] [PubMed]
- 55.Sweller J. Cognitive load theory and educational technology. Educ Tech Res Dev. 2020;68(1):1–16. 10.1007/s11423-019-09701-3. [DOI] [Google Scholar]
- 56.Mayer RE. Cognitive theory of multimedia learning. Cambridge Handb Multimedia Learn. 2005;41(1):31–48. [Google Scholar]
- 57.De Choudhury M, Pendse SR, Kumar N. Benefits and harms of large language models in digital mental health. arXiv preprint arXiv: 2311.14693, 2023.
- 58.Fernández-Luque L, Bau T. Health and social media: perfect storm of information. Healthc Inf Res. 2015;21(2):67–73. 10.4258/hir.2015.21.2.67. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 59.Van Merriënboer JJG, Sweller J. Cognitive load theory in health professional education: design principles and strategies. Med Educ. 2010;44(1):85–93. 10.1111/j.1365-2923.2009.03498.x. [DOI] [PubMed] [Google Scholar]
- 60.Abu-Rasheed H, Weber C, Fathi M. Knowledge graphs as context sources for llm-based explanations of learning recommendations. 2024 IEEE Global Engineering Education Conference (EDUCON). IEEE; 2024.
- 61.Zhang X, et al. Large language models as evaluators for recommendation explanations. Proceedings of the 18th ACM Conference on Recommender Systems. 2024.
- 62.Li Y, et al. Personal llm agents: insights and survey about the capability, efficiency and security. arXiv preprint arXiv: 2401.05459, 2024.
- 63.Calvillo J, Román I, Roa LM. How technology is empowering patients? A literature review. Health Expectations. 2015;18(5):643–52. 10.1111/hex.12089. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 64.Venkatesh V, Thong JY, Xu X. Unified theory of acceptance and use of technology: a synthesis and the road ahead. JAIS. 2016;17(5):328–76. 10.17705/1jais.00428. [DOI] [Google Scholar]
- 65.Faul F, Erdfelder E, Buchner A, Lang A-G. Statistical power analyses using G* power 3.1: tests for correlation and regression analyses. Behav Res Methods. 2009;41(4):1149–60. 10.3758/BRM.41.4.1149. [DOI] [PubMed] [Google Scholar]
- 66.Whitaker C, Stevelink S, Fear N. The use of Facebook in recruiting participants for health research purposes: a systematic review. J Med Internet Res. 2017;19(8):e290. 10.2196/jmir.7071. [DOI] [PMC free article] [PubMed]
- 67.Archibald MM, Munce SE. Challenges and strategies in the recruitment of participants for qualitative research. Univ Alta Health Sci J. 2015;11(1):34–37. [Google Scholar]
- 68.Zhang Z, Yan C, Mesa DA, Sun J, Malin BA. Ensuring electronic medical record simulation through better training, modeling, and evaluation. J Am Med Inf Assoc. 2020;27(1):99–108. 10.1093/jamia/ocz161. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 69.Bodenreider O. The unified medical language system (UMLS): integrating biomedical terminology. Nucleic Acids Res. 2004;32(90001):267D–70. 10.1093/nar/gkh061. [DOI] [PMC free article] [PubMed]
- 70.Thirunavukarasu AJ, et al. Large language models in medicine. Nat Med. 2023;1–11. [DOI] [PubMed]
- 71.Liu C-L, Ho C-T, Wu T-C. Custom GPTs enhancing performance and evidence compared with GPT-3.5, GPT-4, and GPT-4o? A study on the emergency medicine specialist examination. In: Healthcare. MDPI; 2024. [DOI] [PMC free article] [PubMed]
- 72.Ayers JW, Poliak A, Dredze M, Leas EC, Zhu Z, Kelley JB, et al. Comparing physician and artificial intelligence chatbot responses to patient questions posted to a public social media forum. JAMA Intern Med. 2023;183(6):589. 10.1001/jamainternmed.2023.1838. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 73.Shahriar S, Lund BD, Mannuru NR, Arshad MA, Hayawi K, Bevara RVK, et al. Putting gpt-4o to the sword: a comprehensive evaluation of language, vision, speech, and multimodal proficiency. Appl Sci. 2024;14(17):7782. 10.3390/app14177782. [DOI] [Google Scholar]
- 74.Wang L, Wan Z, Ni C, Song Q, Li Y, Clayton E, et al. Applications and concerns of ChatGPT and other conversational Large language models in health care: systematic review. J Med Internet Res. 2024;26:e22769. 10.2196/22769. [DOI] [PMC free article] [PubMed]
- 75.Kung TH, Cheatham M, Medenilla A, Sillos C, De Leon L, Elepaño C, et al. Performance of ChatGPT on USMLE: potential for AI-assisted medical education using large language models. PLoS Digit Health. 2023;2(2):e0000198. 10.1371/journal.pdig.0000198. [DOI] [PMC free article] [PubMed]
- 76.Amugongo LM, et al. Retrieval augmented generation for Large language models in healthcare: a systematic review. 2024. [DOI] [PMC free article] [PubMed]
- 77.Berkman ND, et al. Health literacy interventions and outcomes: an updated systematic review. Evidence report/Technol Assess. 2011;199:1–941. [PMC free article] [PubMed] [Google Scholar]
- 78.Demiris G, Iribarren SJ, Sward K, Lee S, Yang R. Patient generated health data use in clinical practice: a systematic review. Nurs Outlook. 2019;67(4):311–30. 10.1016/j.outlook.2019.04.005. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 79.Rego C, Montague E. The impact of feedback modalities and the influence of cognitive load on interpersonal communication in nonclinical settings: experimental study design. JMIR Hum Factors. 2023;10(1):e49675. 10.2196/49675. [DOI] [PMC free article] [PubMed]
- 80.Shepherd J, Waller A, Sanson-Fisher R, Clark K, Ball J. Knowledge of, and participation in, advance care planning: a cross-sectional study of acute and critical care nurses’ perceptions. Int J Nurs Stud. 2018;86:74–81. 10.1016/j.ijnurstu.2018.06.005. [DOI] [PubMed] [Google Scholar]
- 81.Saha S, Beach MC. Impact of physician race on patient decision-making and ratings of physicians: a randomized experiment using video vignettes. J GEN Intern Med. 2020;35(4):1084–91. 10.1007/s11606-020-05646-z. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 82.Choudhury A, Asan O. Role of artificial intelligence in patient safety outcomes: systematic literature review. JMIR Med Inf. 2020;8(7):e18599. 10.2196/18599. [DOI] [PMC free article] [PubMed]
- 83.Smith KL, Martini J. Patient-provider communication and interactions, in chronic illness care: principles and practice. Springer. p. 195–207. 2023.
- 84.Cegala DJ, Coleman MT, Turner JW. The development and partial assessment of the medical communication competence scale. Health Commun. 1998;10(3):261–88. 10.1207/s15327027hc1003_5. [DOI] [PubMed] [Google Scholar]
- 85.McConnell M, Hou C, Panju M, Panju A, Azzam K. Does testing enhance learning in continuing medical education? Can Med Ed J. 2018;9(3):e83. 10.36834/cmej.42236. [DOI] [PMC free article] [PubMed]
- 86.Murphy PW, et al. Rapid estimate of adult literacy in medicine (REALM): a quick reading test for patients. J Read. 1993;37(2):124–30. [Google Scholar]
- 87.Meakin R. The ‘Medical interview satisfaction scale’ (MISS-21) adapted for British general practice. Fam Pract. 2002;19(3):257–63. 10.1093/fampra/19.3.257. [DOI] [PubMed] [Google Scholar]
- 88.Hart SG. NASA-task load index (NASA-TLX); 20 years later. Proceedings of the human factors and ergonomics society annual meeting. Los Angeles, CA: Sage publications Sage CA; 2006.
- 89.Johnson MO, Rose CD, Dilworth SE, Neilands TB. Advances in the conceptualization and measurement of health care empowerment: development and validation of the health care empowerment inventory. PLoS One. 2012;7(9):e45692. 10.1371/journal.pone.0045692. [DOI] [PMC free article] [PubMed]
- 90.O’Connor AM. Validation of a decisional conflict scale. Med Decis Mak. 1995;15(1):25–30. 10.1177/0272989X9501500105. [DOI] [PubMed] [Google Scholar]
- 91.Hibbard JH, Mahoney ER, Stockard J, Tusler M. Development and testing of a short form of the patient activation measure. Health Serv Res. 2005;40(6p1):1918–30. 10.1111/j.1475-6773.2005.00438.x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 92.Spielberger CD, et al. The state-trait anxiety inventory. Revista Interamericana de Psicologia/Interamerican J Phychol. 1971;5(3 & 4).
- 93.Kvaal K, Ulstein I, Nordhus IH, Engedal K. The Spielberger state-trait anxiety inventory (STAI): the state scale in detecting mental disorders in geriatric patients. Int J Geriat Psychiatry. 2005;20(7):629–34. 10.1002/gps.1330. [DOI] [PubMed] [Google Scholar]
- 94.Appelman A, Sundar SS. Measuring message credibility: construction and validation of an exclusive scale. J Mass Commun Quart. 2016;93(1):59–79. 10.1177/1077699015606057. [DOI] [Google Scholar]
- 95.Montague E. Validation of a trust in medical technology instrument. Appl Ergon. 2010;41(6):812–21. 10.1016/j.apergo.2010.01.009. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 96.Hayes AF. Partial, conditional, and moderated moderated mediation: quantification, inference, and interpretation. Commun Monogr. 2018;85(1):4–40. 10.1080/03637751.2017.1352100. [DOI] [Google Scholar]
- 97.Wilson EV. Modeling patients’ acceptance of provider-delivered E-health. J Am Med Inf Assoc. 2004;11(4):241–48. 10.1197/jamia.M1475. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 98.Lightfoot CJ, Wilkinson TJ, Sohansoha GK, Gillies CL, Vadaszy N, Ford EC, et al. The effects of a digital health intervention on patient activation in chronic kidney disease. NPJ Digit Med. 2024;7(1):318. 10.1038/s41746-024-01296-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 99.Corbin JM, Strauss A. Grounded theory research: procedures, canons, and evaluative criteria. Qual Sociol. 1990;13(1):3–21. 10.1007/BF00988593. [DOI] [Google Scholar]
- 100.Charmaz K. Constructing grounded theory (introducing qualitative methods series). Constr. Grounded Theory. 2014.
- 101.Heath H, Cowley S. Developing a grounded theory approach: a comparison of Glaser and Strauss. Int J Nurs Stud. 2004;41(2):141–50. 10.1016/S0020-7489(03)00113-5. [DOI] [PubMed] [Google Scholar]
- 102.Creswell JW, Miller DL. Determining validity in qualitative inquiry. Theory Pract. 2000;39(3):124–30. 10.1207/s15430421tip3903_2. [DOI] [Google Scholar]
- 103.Lincoln YS, Guba EG. In: Naturalistic inquiry. Newbury Park, CA: Sage Publications; 1985.
- 104.Skulmowski A, Xu KM. Understanding cognitive load in digital and online learning: a new perspective on extraneous cognitive load. Educ Psychol Rev. 2022;34(1):171–96. 10.1007/s10648-021-09624-7. [DOI] [Google Scholar]
- 105.Wang R, Zhou C, Wu Y, Sun M, Yang L, Ye X, et al. Patient empowerment and self-management behaviour of chronic disease patients: a moderated mediation model of self-efficacy and health locus of control. J Adv Nurs. 2022;78(4):1055–65. 10.1111/jan.15077. [DOI] [PubMed] [Google Scholar]
- 106.Esmaeilzadeh P, Hassanein K, Head M. AI-human interactions in healthcare: exploring users’ post-adoption behaviors of AI mental health chatbots. Behaviour Inf Technol. 2025;1–29. 10.1080/0144929X.2025.2477754. [DOI]
- 107.Calisto FM, Santiago C, Nunes N, Nascimento JC. BreastScreening-AI: evaluating medical intelligent agents for human-AI interactions. Artif Intell Med. 2022;127:102285. 10.1016/j.artmed.2022.102285. [DOI] [PubMed] [Google Scholar]
- 108.Ghaffar Nia N, Kaplanoglu E, Nasab A. Evaluation of artificial intelligence techniques in disease diagnosis and prediction. Discov Artif Intell. 2023;3(1):5. 10.1007/s44163-023-00049-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 109.Murtaza M, Cheng C-T, Albahlal BM, Muslam MMA, Raza MS. The impact of LLM chatbots on learning outcomes in advanced driver assistance systems education. Sci Rep. 2025;15(1):7260. 10.1038/s41598-025-91330-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 110.Mendel T, Singh N, Mann DM, Wiesenfeld B, Nov O. Laypeople’s use of and attitudes toward Large language models and search engines for health queries: survey study. J Med Internet Res. 2025;27:e64290. 10.2196/64290. [DOI] [PMC free article] [PubMed]
- 111.Johnson D, et al. Assessing the accuracy and reliability of AI-generated medical responses: an evaluation of the chat-GPT model. Res Square. 2023;rs. 3. rs-2566942.
- 112.Choudhury A, Chaudhry Z. Large language models and user trust: consequence of self-referential learning loop and the deskilling of health care professionals. J Med Internet Res. 2024;26:e56764. 10.2196/56764. [DOI] [PMC free article] [PubMed]
- 113.Longoni C, Bonezzi A, Morewedge CK. Resistance to medical artificial intelligence. J Consum Res. 2019;46(4):629–50. 10.1093/jcr/ucz013. [DOI] [Google Scholar]
- 114.Lankton NK, McKnight DH, Tripp J. Technology, humanness, and trust: rethinking trust in technology. JAIS. 2015;16(10):880–918. 10.17705/1jais.00411. [DOI] [Google Scholar]
- 115.Ongena YP, Haan M, Yakar D, Kwee TC. Patients’ views on the implementation of artificial intelligence in radiology: development and validation of a standardized questionnaire. Eur Radiol. 2020;30(2):1033–40. 10.1007/s00330-019-06486-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 116.Tschandl P, Rinner C, Apalla Z, Argenziano G, Codella N, Halpern A, et al. Human–computer collaboration for skin cancer recognition. Nat Med. 2020;26(8):1229–34. 10.1038/s41591-020-0942-0. [DOI] [PubMed] [Google Scholar]
- 117.Li W, Liu X. Anxiety about artificial intelligence from patient and doctor-physician. Patient Educ Couns. 2025;133:108619. 10.1016/j.pec.2024.108619. [DOI] [PubMed] [Google Scholar]
- 118.Bedi S, et al. Testing and evaluation of health care applications of large language models: a systematic review. JAMA. 2024. [DOI] [PMC free article] [PubMed]
- 119.Giani U, Brascio G, Bruzzese D, Garzillo C, Vigilante S. Emotional and cognitive information processing in web-based medical education. J Biomed Inf. 2007;40(3):332–42. 10.1016/j.jbi.2006.11.004. [DOI] [PubMed] [Google Scholar]
- 120.Power TE, Swartzman LC, Robinson JW. Cognitive-emotional decision making (CEDM): a framework of patient medical decision making. Patient Educ Couns. 2011;83(2):163–69. 10.1016/j.pec.2010.05.021. [DOI] [PubMed] [Google Scholar]
- 121.Schrills T, Franke T. How do users experience traceability of AI systems? Examining subjective information processing awareness in automated insulin delivery (AID) systems. ACM Trans. Interact. Intell. Syst. 2023;13(4):1–34. 10.1145/3588594. [DOI] [Google Scholar]
- 122.Dave DM, Mandvikar S. Augmented intelligence: human-AI collaboration in the era of digital transformation. IJEAST. 2023;8(6):24–33. 10.33564/IJEAST.2023.v08i06.003. [DOI] [Google Scholar]
- 123.Nasarian E, Alizadehsani R, Acharya UR, Tsui K-L. Designing interpretable ML system to enhance trust in healthcare: a systematic review to proposed responsible clinician-AI-collaboration framework. Inf Fusion. 2024;108:102412. 10.1016/j.inffus.2024.102412. [DOI] [Google Scholar]
- 124.Sezgin E. Artificial intelligence in healthcare: complementing, not replacing, doctors and healthcare providers. Digit Health. 2023;9:20552076231186520. 10.1177/20552076231186520. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 125.Li B, et al. A comparative analysis of information gathering by chatbots, questionnaires, and Humans in clinical pre-consultation. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. 2025.
- 126.Zhou J, et al. Artificial intelligence-driven transformative applications in disease diagnosis technology. Med Rev. 2025;. [DOI] [PMC free article] [PubMed]
- 127.Yasaka K, Abe O. Deep learning and artificial intelligence in radiology: current applications and future directions. PLoS Med. 2018;15(11):e1002707. 10.1371/journal.pmed.1002707. [DOI] [PMC free article] [PubMed]
- 128.Tu T, et al. Towards conversational diagnostic AI. arXiv preprint arXiv: 2401.05654, 2024.
- 129.Robertson C, Woods A, Bergstrand K, Findley J, Balser C, Slepian MJ, et al. Diverse patients’ attitudes towards artificial intelligence (AI) in diagnosis. PLoS Digit Health. 2023;2(5):e0000237. 10.1371/journal.pdig.0000237. [DOI] [PMC free article] [PubMed]
- 130.Berdahl CT, Baker L, Mann S, Osoba O, Girosi F. Strategies to improve the impact of artificial intelligence on health equity: scoping review. Jmir Ai. 2023;2:e42936. 10.2196/42936. [DOI] [PMC free article] [PubMed]
- 131.Tappen RM, Cooley ME, Luckmann R, Panday S. Digital health information disparities in older adults: a mixed methods study. J Racial Ethnic Health Disparities. 2022;9(1):82–92. 10.1007/s40615-020-00931-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 132.Hall AK, Bernhardt JM, Dodd V, Vollrath MW. The digital health divide: evaluating online health information access and use among older adults. Health Educ Behav. 2015;42(2):202–09. 10.1177/1090198114547815. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 133.Dwivedi YK, Rana NP, Jeyaraj A, Clement M, Williams MD. Re-examining the unified theory of acceptance and use of technology (UTAUT): towards a revised theoretical model. Inf Syst Front. 2019;21(3):719–34. 10.1007/s10796-017-9774-y. [DOI] [Google Scholar]
- 134.Maruping LM, Bala H, Venkatesh V, Brown SA. Going beyond intention: integrating behavioral expectation into the unified theory of acceptance and use of technology. Asso Info Sci Tech. 2017;68(3):623–37. 10.1002/asi.23699. [DOI] [Google Scholar]
- 135.Granić A, Marangunić N. Technology acceptance model in educational context: a systematic literature review. Brit J Educ Tech. 2019;50(5):2572–93. 10.1111/bjet.12864. [DOI] [Google Scholar]
- 136.Esmaeilzadeh P, Mirzaei T, Maddah M. The effects of data entry structure on patients’ perceptions of information quality in health information exchange (HIE). Int J Med Inf. 2020;135:104058. 10.1016/j.ijmedinf.2019.104058. [DOI] [PubMed] [Google Scholar]
- 137.Pan X. Technology acceptance, technological self-efficacy, and attitude toward technology-based self-directed learning: learning motivation as a mediator. Front Psychol. 2020;11:564294. 10.3389/fpsyg.2020.564294. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 138.Liu PL, Ye JF. Mobile patient-provider communication and lifestyle improvement: examining the role of mobile technology identity and health empowerment. Health Commun. 2025;40(8):1415–27. 10.1080/10410236.2024.2402160. [DOI] [PubMed] [Google Scholar]
- 139.Wies B, Landers C, Ienca M. Digital mental health for young people: a scoping review of ethical promises and challenges. Front Digit Health. 2021;3:697072. 10.3389/fdgth.2021.697072. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 140.Petrakaki D, Hilberg E, Waring J. The cultivation of digital health citizenship. Soc Sci Med. 2021;270:113675. 10.1016/j.socscimed.2021.113675. [DOI] [PubMed] [Google Scholar]
- 141.Zamani H, Trippas JR, Dalton J, Radlinski F. Conversational information seeking. Found Trends® Inf Retr. 2023;17(3–4):244–456. 10.1561/1500000081. [DOI] [Google Scholar]
- 142.Rose-Davis B, Van Woensel W, Raza Abidi S, Stringer E, Sibte Raza Abidi S. Semantic knowledge modeling and evaluation of argument theory to develop dialogue based patient education systems for chronic disease self-management. Int J Med Inf. 2022;160:104693. 10.1016/j.ijmedinf.2022.104693. [DOI] [PubMed] [Google Scholar]
- 143.Yue C, Kim J, Ogawa R, Stark E, Kim S. Applying the cognitive theory of multimedia learning: an analysis of medical animations. Med Educ. 2013;47(4):375–87. 10.1111/medu.12090. [DOI] [PubMed] [Google Scholar]
- 144.Mayer RE. Using multimedia for e-learning. J Comput Assisted Learn. 2017;33(5):403–23. 10.1111/jcal.12197. [DOI] [Google Scholar]
- 145.Riefle L, et al. Adaptive conversational agents: exploring the effect of individualized design on user experience. ECIS. 2023.
- 146.Song C, Shin S-Y, Shin K-S. Implementing the dynamic feedback-driven learning optimization framework: a machine learning approach to personalize educational pathways. Appl Sci. 2024;14(2):916. 10.3390/app14020916. [DOI] [Google Scholar]
- 147.Rouleau G, Wu K, Ramamoorthi K, Boxall C, Liu RH, Maloney S, et al. Mapping theories, models, and frameworks to evaluate digital health interventions: scoping review. J Med Internet Res. 2024;26:e51098. 10.2196/51098. [DOI] [PMC free article] [PubMed]
- 148.Godinho MA, Ashraf MM, Narasimhan P, Liaw S-T. Understanding the convergence of social enterprise, digital health, and citizen engagement for co-producing integrated person-centred health services: a critical review and theoretical framework. Int J Med Inf. 2023;178:105174. 10.1016/j.ijmedinf.2023.105174. [DOI] [PubMed] [Google Scholar]
- 149.Dodds S, Russell–Bennett R, Chen T, Oertzen A-S, Salvador-Carulla L, Hung Y-C. Blended human-technology service realities in healthcare. JSTP. 2022;32(1):75–99. 10.1108/JSTP-12-2020-0285. [DOI] [Google Scholar]
- 150.Zhang R, Zhang Z, Wang D, Liu Z. Editorial: responsible AI in healthcare: opportunities, challenges, and best practices. Front Comput Sci. 2023;5:1265902. 10.3389/fcomp.2023.1265902. [DOI] [Google Scholar]
- 151.Abràmoff MD, Tarver ME, Loyo-Berrios N, Trujillo S, Char D, Obermeyer Z, et al. Considerations for addressing bias in artificial intelligence for health equity. NPJ Digit Med. 2023;6(1):170. 10.1038/s41746-023-00913-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 152.Moy S, Irannejad M, Jeanneret Manning S, Farahani M, Ahmed Y, Gao E, et al. Patient perspectives on the use of artificial intelligence in health care: a scoping review. J Patient Cent Res Rev. 2024;11(1):51. 10.17294/2330-0698.2029. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
A significant portion of the data analyzed in this study is included in this published article and three Appendices. The data used and analyzed during the current study are available from the corresponding author upon reasonable request.