Application of artificial intelligence to measure and predict patient values and preferences: a scoping review

Abstract

Patients’ voices are often difficult to capture directly in healthcare decisions. This scoping review examines how artificial intelligence (AI) has been applied to measure and predict patient values and preferences, aiming to evaluate its potential to generate reliable, patient-centered evidence, identify opportunities and challenges, and explore AI tools in literature reviews. Analyzing 67 studies, we summarize how AI processes diverse data sources such as social media, clinical records, and patient surveys to extract population- and individual-based patient values and preferences. Researchers have applied AI for efficient data preprocessing, extraction, analysis, integration, and modeling. Despite promising validation results (e.g., >80% accuracy in data preprocessing), key challenges remain, such as data quality issues, lack of real-world validation, and ethical concerns. This review underscores the potential of AI in patient values and preferences research and calls for greater transparency and real-world implementation to better align healthcare delivery with patient needs.

Introduction

Patient values and preferences are the unique understandings, priorities, concerns, expectations and life circumstances of each patient¹. Values reflect patients’ attitudes and perceptions about healthcare options, and preferences are their preferred choices after accounting for their values². Recognizing the critical role of patient values and preferences in aligning healthcare with individual needs, regulatory agencies such as the US Food and Drug Administration^3,4 and the European Medicines Agency strongly^5,6 advocate for patient-centric healthcare⁵. Respecting patient values and preferences is ethically aligned with autonomy and dignity⁷, socially reflects a humanistic approach to care⁸, and clinically informs outcome selection and benefit–risk trade-offs^4,5,9.

The measurement of patient values and preferences involves qualitative methods (e.g., interviews, focus groups) and quantitative methods (e.g., discrete choice experiments, DCEs)^10,11. Existing methods face limitations: qualitative methods often suffer from small samples and limited reproducibility, while quantitative methods are time-intensive. Both methods involve high cognitive demands and struggle to integrate diverse data sources^12–14. In addition, DCE, the widely used quantitative method, relies on hypothetical scenarios and thus often fail to capture patients’ real-world preferences¹⁵. Researchers have also explored patient values and preferences prediction, such as the Patient Preference Predictor (PPP), particularly for patients unable to make decision¹⁶. Traditional statistical prediction models using static evidence face concerns related to autonomy, fairness, trust, and adaptability, and struggle with reproducibility across different populations^16,17.

Artificial Intelligence (AI)¹⁸ technologies in patient-centric healthcare present opportunities for more efficient and accurate measurement and prediction of patient values and preferences^19,20. For example, Natural Language Processing (NLP)¹⁸ can automatically and rapidly analyze unstructured data from social media, clinical records, and other sources. And thus, facilitate the large-scale identification of individual concerns, expectations, and value-laden expressions that underpin patients’ preferences. By identifying emotional cues and thematic patterns embedded in patients’ natural language, sentiment analysis and topic modeling can help capture real-world patient values and preferences more effectively^21,22. Machine learning (ML) and deep learning (DL)¹⁸, particularly large language models (LLMs), can process high-dimensional, multi-source data to identify complex preference patterns, adapt to evolving patient information, and enhance context-aware, scalable modeling of nuanced patient values and preferences^22,23. By automating data processing and applying standardized algorithms, AI technologies can enhance the scalability and reproducibility of patient values and preferences measurement and prediction^21–23. Despite its potential to personalize care and support decision-making—especially when patients are unable to express their preferences—AI faces major challenges in interpretability (e.g., limited transparency in how preference predictions are generated), bias (e.g., underrepresentation of older adults or underserved populations in training data), fairness (e.g., inconsistent performance when applied across different sociocultural groups), and integration into clinical workflows (e.g., lack of validation in real-world settings where preference-sensitive decisions are made)^19,20

Clarifying the current landscape of how AI has been used to measure and predict patient values and preferences is essential to guide future efforts. Accordingly, this review aims to: (1) characterize how AI technologies have been applied to measure and predict patient values and preferences; (2) identify opportunities and challenges in integrating AI into patient values and preferences studies; and (3) reflect on practical insights gained from exploratory use of AI tools in screening and data extraction.

Results

Study characteristics and methodological quality

The search strategy yielded 12466 records, with 67 studies included in this review (Fig. 1). All ultimately serves patient-centered healthcare, with most focusing on clinical applications such as disease management and treatment (23/67), health service optimization (19/67), and medical decision-making (17/67). Table 1 summarizes their basic characteristics, with detailed information provided in the Supplementary Table 2. Methodological quality was mostly moderate (33/67) or high (32/67), with one study rated low; one protocol lacked an appropriate tool and was not assessed. Supplementary Table 3 provides detailed results of methodological quality assessment.

Fig. 1. Study selection.

Flow diagram showing the selection process of included studies in accordance with PRISMA-ScR guidelines.

Table. 1.

Summary of basic characteristics of the included 67 studies

Study characteristic	Category	Number of studies (n = 67)	Proportion (%)
Publication year	2024	26	39%
	2023	12	18%
	2022	4	6%
	2021	6	9%
	2020	6	9%
	2019	5	7%
	2018	1	1%
	2017	3	4%
	2016	0	0%
	2015	2	3%
	2014	2	3%
Country (Six were multinational collaborations)	United States	25	37%
	China	12	18%
	Canada	6	9%
	United Kingdom	4	6%
	Germany	4	6%
	Other 14 countries and multi-region collaborations	1–3	1–4%
Study type	Peer-reviewed journal article	45	87%
	Original research	24	36%
	Opinion/Commentary	10	15%
	Review	5	7%
	Protocol	1	1%
	Methodological/ conceptual study	4	6%
	Original research (abstract)	1	1%
	Conference paper (full-text)	15	22%
	Conference paper (abstract)	4	6%
	Book Chapter	2	3%
	Guideline	1	1%
Healthcare context	Disease management and treatment (e.g., chronic conditions, palliative care)	23	34%
	Health service and process (e.g., physician or hospital recommendation)	19	28%
	Medical decision-making (including clinical decision support and shared decision-making)	17	25%
	Ethical considerations	2	3%
	Other (e.g., clinical practice guideline development, drug development)	6	9%

Based on group discussions using the Systematized Nomenclature of Medicine – Clinical Terms⁷⁵, we refined healthcare context into five categories.

Using population-based or individual based data from social media, electronic health records, and patient-reported sources (Fig. 2b), researchers have applied AI to measure and predict patient values and preferences in disease management and treatment, medical service optimization, and medical decision support (including clinical decision support and shared decision-making). Specifically, researchers have applied AI in three key areas: measuring patient values and preferences (32 studies), synthesizing primary studies (3 studies), and predicting patient values and preferences (32 studies). These applications involved multiple steps such as data preprocessing (e.g., cleaning and screening), extraction, analysis, integration, and modeling (Fig. 2). Table 2 summarizes the AI technologies and algorithms applied in 67 studies and their respective roles.

Fig. 2. Comprehensive overview of artificial intelligence applications in patient values and preferences—measurement, evidence synthesis, and prediction.

Figure 2 presents the comprehensive overview of AI applications in patient values and preferences, including the three key types involved (a), the characteristics of the data used in these studies (b–d), the mapping of AI applications across different research stages (e), and the metrics and methods for AI performance evaluation (f). AI artificial intelligence.

Table. 2.

Summary of AI technologies and algorithms applied in patient values and preferences¹⁸

A single study may use multiple AI methods; total exceeds 67. Color shading indicates the number of studies using each technology, with darker shades reflecting higher frequencies.

AI artificial intelligence.

AI in measurement of patient values and preferences

Thirty-two studies explored the use of AI in measuring patient values and preferences, employing two main settings. Supplementary Table 2 presents detailed information.

One setting involves directly applying AI to process large-scale real-world data sources—such as health-related websites (e.g., Healthgrades, Drugs.com), social media platforms (e.g., Reddit disease subreddits, Facebook support groups), and electronic health records—to analyze and extract patient values and preferences (25/32), with most of the data being population-based (23/25). For example, Lösch et al. used NLP and ML techniques to identify patient values, preferences and experiential knowledge regarding COVID-19 vaccination from over 230,000 Facebook comments and two national health databases in the Netherlands²². After filtering for highly subjective and emotionally charged texts, they trained an ML classifier to detect narratives containing experiential information. Structural topic modeling was then applied to extract and categorize recurring themes related to patient experiences and concerns. This approach enabled the identification of specific value-laden issues—such as unmet needs for vaccination in cross-border caregiving or perceived inconsistencies in medical prioritization—that may not have surfaced through traditional guideline development methods. Liao et al. applied traditional ML models (e.g., logistic regression, Light Gradient Boosting Machine) and a LLM (bidirectional encoder representations from transformers) to classify and score patient online reviews from two Chinese hospital review platforms²³. They extracted preference-related attributes (e.g., service attitude, facility environment), quantified sentiments using fuzzy logic, and integrated them with structured hospital indicators to generate personalized hospital rankings aligned with patient values.

Another setting involves using AI to process existing patient values and preferences data from patient survey—such as interviews and questionnaires—into structured formats for further application (6/32), with most of the data being individual-based (4/6). In this context, AI techniques focused more on standardizing patient narratives and inferring personalized models. For example, van den Broek-Altenburg et al. used NLP with the Moral Foundations Dictionary to extract moral values from transcribed palliative care consultations and identify patterns associated with end-of-life preferences²¹. Latent class analysis revealed distinct moral expression groups linked to patient demographics and spiritual needs. Birch et al. proposed using a questionnaire-based risk–value profiling method to elicit patients’ preferences regarding false positives, false negatives, and treatment trade-offs. This information could then inform decision thresholds in AI diagnostic tools, enabling tailored recommendations aligned with each patient’s values¹⁹. However, we did not identify any studies that directly applied AI to enhance patient values and preference surveys; for example, we excluded one that converted DCEs (one of the best methods to quantitatively measure patient values and preferences¹⁵) into an online format without using AI²⁴.

Overall, measurement studies predominantly relied on large-scale, population-based data from social media and online platforms, with few using individual-level survey data. Notably, no study applied AI to enhance established elicitation methods such as discrete choice experiments, underscoring limited methodological innovation

Twenty studies evaluated AI performance in patient values and preferences measurement. For data extraction and analysis (especially classification and labeling), AI demonstrated moderate to high accuracy (7/20, 75–99%)^22,25–29 and F1-scores (3/20, 0·9084–0·935; 1, 0·47)^22,23,26,28. A higher F1-score indicates better balance between correctly identifying relevant results (precision) and capturing all relevant results (recall). Six studies incorporated manual reviews or human-in-the-loop comparisons with experts, although only one provided specific data (the effectiveness of aspect-opinion extraction achieved substantial agreement after expert validation, Fleiss’ Kappa = 0·731)³⁰. Other evaluations included relevance analysis (5/20), robustness (1/20), and usability (1/20), though most did not report specific values.

AI in evidence synthesis of patient values and preferences

Three studies discussed the application of AI in synthesizing the primary studies on the measurement of patient values and preferences (Supplementary Table 2 presents detailed information). Brozek et al. employed an AI tool (Laser AI³¹) to combine quantitative patient values and preferences studies on allergic rhinitis (i.e., utility studies, discrete choice experiments, and ranking and rating studies)³². Laser AI utilized NLP to automatically analyze and screen titles and abstracts, identifying relevant content for text filtering, classification, and information extraction. It utilized ML for deduplication, prioritized studies based on relevance and quality, and trained models to automatically assess the risk of bias. Additionally, Laser AI applied DL models to identify complex patterns and relationships within the literature.

Nilsen et al. proposed a conceptual perspective that how AI could help overcome challenges in incorporating patient values and preferences into evidence-based healthcare³³. For example, AI could support the incorporation of patient values and preferences through technologies such as speech recognition for automatically recording doctor-patient conversations and generative AI to assist communication, increasing patient engagement. Automated tools (e.g., chronic disease management devices) could empower patients to actively participate in decision-making, while advanced AI algorithms could analyze historical health data and patient values and preferences to deliver personalized medical recommendations. However, as a conceptual perspective, the study did not detail how these technologies would be implemented or used by patients and clinicians in practice.

Braithwaite et al. proposed that emphasized the potential of AI in integrating patient values and preferences, optimizing medical decision-making, and advancing personalized treatment³⁴. They proposed that building learning health systems—integrating real-time clinical, patient, and cost data and supported by AI technologies—could better incorporate patient values and preferences into care delivery. Through continuous feedback loops and dynamic data-driven decision-making, such systems could facilitate a shift from physician-centered healthcare models toward more patient-centered and adaptive approaches.

Existing applications of AI in evidence synthesis remain limited to a few proof-of-concept studies, mainly on quantitative data, with no systematic use for qualitative or real-world patient preference evidence.

AI in prediction of patient values and preferences

Thirty-two studies applied AI in predicting patient values and preferences (Supplementary Table 2 presents detailed information), encompassing two major settings.

The first setting focused on directly predicting patients’ treatment decisions, particularly in situations where patients were incapacitated or where decision-making was delegated (11/32). Data sources included population-level data—such as information from public social media or health platforms, historical medical records, and large-scale patient values and preferences surveys. Rid et al. initially proposed the concept of a Patient Preference Predictor (PPP)³⁵ suggesting that demographic key characteristics (e.g., age, income) can predict patient values and preferences. Another data sources are individual-level data, including patients’ prior medical histories, decisions, social media information, and results of the patient values and preferences survey. Building on the PPP concept, Earp et al. expanded to a personalized version (i.e., Personalized Patient Preference Predictor, also called P4), leveraging LLMs to extract individual preferences or values from patient-generated or patient-related data³⁶. Both conceptualizations aim for more accurate prediction of patient values and preferences than human surrogate decision-makers. The research in this area, however, remains mostly theoretical. In a proof-of-concept study, Nolan et al. simulated clinical scenarios for 50 incapacitated patients, combining synthetic preference data (numeric questionnaires, graded textual questionnaires, free-text narratives) with a pre-trained BERT model (a DL model that understands text context bidirectionally, widely used for NLP tasks such as question answering, text classification, and named entity recognition) extract the proposed treatment from patient scenario texts. Then using GPT-3.5, Nolan et al. generated personalized treatment recommendations for the simulated incapacitated patients)³⁷. The BERT model successfully extracted proposed treatments in 88% of cases, and the GPT-3.5-generated recommendations were rated as medically reasonable (mean score: 3·92/5) and moderately aligned with patient values (mean score: 3·54/5). However, adjudicator agreement on value alignment was limited, with full consensus reached in only 25% of cases, suggesting both potential and limitations of current LLMs in capturing nuanced patient preferences.

The second setting focused on analyzing existing patient values and preferences data to identify trends, cluster patient groups, and predict preference patterns for applications in disease management, medical decision-making, and healthcare service optimization (21/32). Data sources similarly included both population- and individual-level data, derived from public platforms, electronic medical/health records, large- or small-sample patient values and preferences surveys, and, additionally, mobile-based healthcare applications capturing individual patient inputs. Notably, very few studies aimed to help patients better understand and reflect on their own preferences. Most AI approaches prioritized predicting preferences to support medical decision-making or medical service optimization, rather than fostering individual self-awareness.

Prediction studies, therefore, emphasized decision support over patient empowerment, and most remained theoretical or proof-of-concept, with minimal external validation in real-world healthcare.

Sixteen studies evaluated AI performance in patient values and preferences prediction. Data analysis and preference prediction showed moderate to high accuracy (9/16, 66–94%)^37–45 and moderate F1-score (3/16, 0·68–0·94)^38,44,46, precision (2/16, 77·73–85%)^40,46, and recall (2/16, 59·8–81%)^40,46,47. Four studies conducted manual assessment: two studies indicated high user satisfaction but did not provide specific data^40,48; and one discussed AI as human surrogates in predicting patient values and preferences⁴⁹.

Limitations and challenges of AI applied to patient values and preferences

Figure 3 summarizes the limitations and challenges of AI in the research of patient values and preferences. In measurement of patient values and preferences, key limitations include lack of data representativeness (e.g., social media analysis is restricted to public platforms, excluding patients who do not use social media) and insufficient integration of clinical data and contextual factors (e.g., social media data lacks linkage with medical records and does not account for demographic and socioeconomic factors). AI also struggles with language and context comprehension (e.g., ambiguity in sentiment analysis, difficulty interpreting metaphors and slang) and dynamic tracking of patient values and preferences (e.g., relying on static data without capturing preference changes over time as treatment progresses), limiting its ability to fully reflect patient needs. In synthesizing primary studies on measurement of patient values and preferences, challenges include data quality issues, integration complexity, and limited validation, indicating difficulties in systematically consolidating patient values and preferences data. In prediction of patient values and preferences, challenges stem from insufficient AI explainability and ethical concerns. The most critical limitation is lack of clinical utility validation, meaning that most AI models have not been prospectively tested in real-world healthcare settings. Furthermore, the opacity of training datasets for LLMs—including the unclear representation of different patient subpopulations and sociocultural groups—raises additional concerns about the generalizability and equity of AI-derived patient values and preferences predictions. These challenges highlight the urgent need for transparent reporting, external validation, and clear regulatory definitions to ensure safe, ethical, and reliable integration of AI technologies into patient-centered healthcare.

Fig. 3.

Mapping the limitations and challenges of AI in patient values and preferences studies. AI artificial intelligence.

Performance of generative AI-assisted data extraction

Compared with the manually validated dataset, ChatGPT 4o demonstrated an omission rate of 9% (i.e., missing or partially extracted information) and an error rate of 2% (i.e., fabricated information not present in the original text; Supplementary Table 2). Most issues occurred when study descriptions were vague or indirect (requiring human interpretation), when terminology was inconsistent, when full-text PDFs were long and complex, or when prompts failed to reflect the diversity of study designs. ChatGPT also occasionally misclassified study types, such as labeling IEEE abstracts as original research. While AI outputs offered partial utility, accuracy depended heavily on prompt design and human oversight.

Discussions

This scoping review systematically searched and investigated how AI is used in measuring and predicting patient values and preferences, both in primary research and through evidence synthesis. The methodological quality of included studies was mostly moderate to high (65/67). AI in measuring patient values and preferences is applied in two settings: one uses AI to extract and analyze population-based values and preferences from large-scale real-world unstructured data sources, such as social media and health-related websites, while the other applies AI to assist in converting semi-structured data from patient values and preference surveys (e.g., interviews or questionnaires) into structured formats. AI-assisted evidence synthesis of patient values and preferences includes systematic reviews of quantitative studies, although AI-assisted synthesis for qualitative studies and other data sources still requires further exploration. In predicting patient preferences, one setting uses AI to assist decision-making for incapacitated patients, utilizing both population-level data (e.g., social media, health platforms, medical records) and individual-level data (e.g., medical histories, patient surveys). The second setting uses AI to analyze patient preference data to identify trends and predict preferences for disease management, medical decision-making, and medical service optimization, drawing on population- and individual-level data from surveys, electronic health records, and mobile healthcare applications. However, few studies focus on helping patients understand their own preferences. AI has automated key stages of patient values and preferences studies, particularly in data processing and analysis, improving efficiency and expanding analytical capabilities. For example, NLP, which relies on ML models, has enhanced the extraction of sentiment, key topics, and semantic information from unstructured data. DL, a specialized subset of ML, has further facilitated preference classification, prediction, and contextual understanding by leveraging multilayered neural networks to capture complex patterns. Despite promising performance in terms of accuracy and consistency (reported in some studies), AI applications face challenges, including limited real-world validation, and inadequate integration of clinical and demographic data.

Compared to traditional methods for measuring patient values and preferences (e.g., interviews or DCE)¹¹, AI-driven methods offer several advantages. First, AI leverages a broader range of data sources and large datasets, providing more comprehensive and extensive information^50,51. For example, Shufeng et al. identified five key public health applications of social media during COVID-19, like investigating public attitudes and predicting cases⁵². Jiageng et al. reviewed publicly available clinical text datasets for medical AI⁵³. Second, AI significantly accelerates data processing, reduces manual screening time, and enhances the ability to capture critical insights⁵⁴. For instance, NLP filters (e.g., sentiment and subjectivity scores) automatically extract experiential knowledge⁵⁵, while automated tokenization, stemming, and stop-word filtering simplify preprocessing³⁹. Beyond patient values and preferences, AI supports broader medical research domains. Ryan et al. reviewed randomized controlled trials of AI algorithms in clinical practice, focusing on deep learning systems for medical imaging⁵⁶. Fusiak et al. explored incorporating patient values and preferences into decision-making models⁵⁷, and Jonathan et al. emphasized that clinical decisions using AI must consider patient values¹⁹, positioning AI as an assistive tool rather than a replacement for human decision-makers²⁰. Third, AI can help uncover latent patient values and preferences that are not explicitly expressed or easily accessible, interpreting nuanced emotions and unspoken priorities or needs⁵⁸. While clinicians and validated tools (e.g., the checklist of nonverbal pain indicators)⁵⁹ are also capable of recognizing such nonverbal cues, AI offers the potential to process these signals at scale and integrate them with large, multimodal datasets, thereby providing complementary insights. For example, a patient might prioritize vaccine safety over the speed of administration²². Topic modeling techniques capture semantic relationships within texts, revealing hidden patterns and insights⁶⁰. Fourth, AI enhances the reproducibility and objectivity of research outcomes, mitigating the biases often present in traditional analysis methods due to researcher subjectivity or inconsistent data quality⁶¹. Specifically, automated tools ensure consistency in data cleaning and analysis, while machine learning models deliver highly consistent algorithmic results, reducing the influence of human intervention⁶².

These strengths resonate with broader theoretical perspectives. From a health psychology perspective⁶³, AI-based analyses of patient narratives help capture attitudes, beliefs, and behavioral intentions, aligning with theories of behavior change. From decision theory, AI’s modeling of trade-offs between benefits and harms reflects principles of expected utility and multi-criteria decision-making⁶⁴. From clinical ethics⁴³, AI highlights tensions between beneficence and respect for autonomy, while offering opportunities to support value-concordant care.

The integration of multiple AI technologies drives a shift from traditional data collection to intelligent analysis, enabling clinical practice guideline development⁶⁵, medical decision-making¹⁹, personalized medicine^36,43, and healthcare services²³, while advancing the transition toward patient-centered care. In terms of clinical applicability, AI methods could be embedded into concrete scenarios. For instance, sentiment analysis has potential to inform shared decision-making processes by clarifying patient values³⁹, while preference prediction models may assist in end-of-life care planning for incapacitated patients³⁵. AI-driven prioritization models could also support treatment allocation decisions, such as determining priority for limited healthcare resources, thereby aligning care delivery with patient values. Such examples illustrate how methodological advances can translate into practical, patient-centered decision support without replacing clinician judgment.

Although AI demonstrates significant potential in patient values and preferences, this review identifies several limitations. First, data from public platforms such as social media often lack direct relevance to specific patient populations and may fail to capture authentic preferences, risking depersonalized inferences that overlook individual variability and potentially erode patient–clinician trust⁶⁶. Second, training datasets frequently lack geographic, ethnic, and socioeconomic diversity, overrepresenting white, urban, and higher-income populations. Such imbalance reinforces algorithmic bias and raises serious equity concerns, particularly for marginalized groups (e.g., rural or indigenous populations) who remain digitally underrepresented. Third, the use of AI in analyzing patient data raises critical concerns on ethics⁶⁷, privacy, and fairness, which may undermine patient trust and limit broader adoption of these technologies. At the same time, some emerging approaches—such as generative models that produce synthetic biomedical data—seek to address these issues by enhancing privacy and reducing bias⁶⁸. Fourth, current AI models lack explainability⁶⁹, resulting in insufficient transparency and diminishing the trust of patients and healthcare providers in AI-driven decisions. While improving explainability is important, future regulatory frameworks should acknowledge the inherent opacity of some AI systems and mandate verifiable external validation to ensure trustworthy use in patient values and preferences studies. Fifth, AI applications in patient values and preferences involve two primary approaches, each with limitations: (1) inferring preferences based on population-level averages among patients with similar backgrounds and contexts, which may overlook individual variability⁷⁰; and (2) in ideal scenarios, constructing digital twins using individual patients’ historical data (e.g., previous treatment decisions, daily preferences) to predict personalized preferences or extracting preferences directly from patient-generated materials. However, this second approach requires comprehensive and reliable personal data, which are often difficult to obtain. Sixth, most existing studies focus on theoretical validation, with limited application in real-world clinical settings. Consistent with our review, Loftus et al. also noted that the application of AI in predicting patient values and preferences remains at a theoretical stage⁷¹. Furthermore, given the early-stage development and limited validation of most included AI models, their predictive capacity may be overstated—especially when evaluated solely through internal metrics without real-world testing⁵⁶. Seventh, research on AI applications in evidence synthesis of patient values and preferences is limited. Francisco et al. reviewed several AI tools for literature reviews (e.g., screening and data extraction)⁶⁸,but their applicability to synthesizing patient values and preferences evidence requires further investigation.

In summary, the application of AI in measuring and predicting patient values and preferences remains primarily auxiliary, optimizing traditional methods, expanding information sources, enhancing efficiency, and strengthening data processing capabilities.

Our scoping review is the first to comprehensively analyze how AI is used in measuring and predicting patient values and preferences, emphasizing how AI can address challenges in obtaining reliable patient values and preferences evidence and create new opportunities to advance patient-centric healthcare. Although formal methodological quality assessment is not required in scoping reviews, we still evaluated the methodological quality of included studies using design-appropriate tools (e.g., QuADS, MI-CLAIM) or the AACODS checklist for grey literature, to enhance rigor and transparency. We incorporated AI tools to support the review process: ASReview was used for screening prioritization, and ChatGPT 4o was tested for exploratory data extraction. All final data were manually extracted and verified, with ChatGPT outputs used solely for comparative evaluation. While performance was acceptable (omission: 9%, error: 2%), this exploratory use highlighted key limitations of proprietary, continuously updated models such as ChatGPT—particularly their lack of reproducibility. This underscores the importance of transparency and version control in future applications of generative AI in evidence synthesis. This review has some limitations. First, we only included studies published in English, and might have missed relevant studies reported in other languages. Second, while we incorporated gray literature and expert opinions to ensure a comprehensive investigation, some studies may have lacked complete information. To mitigate this limitation, we conducted additional manual and backward snowballing⁷² to retrieve as much relevant data as possible.

Methods

We registered this scoping review on the Open Science Framework (osf.io/dft9j) and followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for scoping reviews (PRISMA-ScR)⁷³.

Search strategy and selection criteria

Using PubMed, Embase (Ovid), Scopus, CINAHL, PsycINFO, and IEEE Xplore (from inception through 19 November, 2024), we searched for studies simultaneously focused on “artificial intelligence” and “patient values and preferences”. Supplementary Table 1 provides detailed search terms and strategies for each database. We included English studies of any design, including peer-reviewed journal studies (original research, review, commentary, etc), conference papers, guidelines, and book chapters. We also manually reviewed the references of pertinent publications to identify additional literature (i.e., backward snowballing)⁷².

We imported all retrieved records into EndNote 20.2 (Clarivate, Philadelphia, PA, USA) for deduplication prior to screening. We used ASReview software (version 1.6.3, Utrecht University, the Netherlands)⁷⁴ to assist two independent reviewers (MTY and YL) in screening study titles and abstracts, followed by a manual full-text review of the selected papers. A third reviewer (LNZ) resolved any discrepancies through discussion. This open-source tool uses active learning algorithms to prioritize potentially relevant records. Before initiating the screening process, we manually labeled six relevant and seven irrelevant records as prior knowledge to warm up the model. Reviewers manually screened all records suggested by the algorithm and did not apply an automatic stopping rule. Supplementary Note 1 provides detailed software settings and screening logs.

Data extraction and analysis

Two reviewers (TH and YL) independently extracted data using a pilot-tested data abstraction form following a calibration exercise. A third reviewer (LNZ) resolved any discrepancies through discussion. To explore the feasibility of generative AI for patient values and preferences data extraction, a fourth reviewer (MTY) conducted a parallel, exploratory data extraction using ChatGPT 4o (OpenAI, San Francisco, CA, USA), following prompt refinement based on ten pilot studies. Reviewers then cross-verified all AI-generated outputs against the manually validated dataset and classified inconsistencies as omissions (i.e., missing or partially extracted information) or errors (i.e., fabricated or incorrect content). We based the final dataset used for analysis solely on the manually validated version.

Data abstraction focused on five aspects: (1) study characteristics, including publication year, country, study type and design, and healthcare context; (2) AI data modalities, including data source (e.g., social media such as Reddit, Facebook, health-related websites, electronic medical records, and patient values and preferences studies), type, and level; (3) the stage in which the researchers applied AI (e.g., data preprocessing and extraction) and technologies applied (e.g., NLP, ML); (4) the evaluation of AI performance (e.g., the accuracy and precision of AI in measuring or predicting patient values and preferences); (5) limitations and challenges of AI in patient values and preferences studies, as explicitly reported or discussed by the study authors.

To systematically summarize the applications of AI in patient values and preferences across the included studies, we standardized the categorization and analysis of extracted data. Based on group discussions using the Systematized Nomenclature of Medicine – Clinical Terms⁷⁵, we refined healthcare context into five categories—disease management and treatment, health service and process, medical decision-making, ethical considerations, and other (e.g., clinical practice guideline development, drug development). Following the International Organization for Standardization 2382 standard⁷⁶, we classified patient values and preferences data sources into structured, semi-structured, and unstructured formats. According to the Health Level Seven Fast Healthcare Interoperability Resources standard⁷⁷, we classified data levels as population-based, individual-based, or both. Using the AI lifecycle framework⁷⁸, we divided AI applications in measuring and predicting patient values and preferences into five stages: data preprocessing (e.g., cleaning, screening), extraction, analysis, integration, and modeling. We excluded data collection since the included studies did not report whether AI technologies were used for this purpose or specify the techniques applied. Based on the Association for Computing Machinery Computing Classification System⁷⁹, we further classified AI technologies and algorithms, covering methods such as traditional ML (a subfield of AI enabling systems to learn from data), DL (an important branch of ML), LLM (a type of DL model), NLP (an AI application area focused on human language), along with other AI technologies and specific algorithms. Although some classified technologies span different hierarchical levels within AI, and even overlap in their relationships, we highlighted the major technologies most commonly applied in patient values and preferences studies. Meanwhile, we summarized AI performance evaluation metrics and results into quantitative and qualitative assessments. Supplementary Note 2 provides detailed classification standards and definitions. Lastly, we conducted a thematic analysis of the reported limitations and challenges to identify key themes and descriptively synthesize them into structured categories for clearer interpretation.

Methodological quality assessments

Two reviewers (MTY and YL)received training, calibrated their judgments, and independently assessed all studies in duplicate. For peer-reviewed journal articles or peer-reviewed conference full texts, we used design-specific tools (e.g., the Quality Assessment with Diverse Studies (QuADS) for mixed- or multi-method studies⁸⁰, the Minimum Information about Clinical Artificial Intelligence Modeling (MI-CLAIM) checklist for AI algorithms in medicine⁸¹). For grey literature, we used the Authority, Accuracy, Coverage, Objectivity, Date, and Significance (AACODS) checklist⁸². The study team defined overall methodological quality on a four-level scale (high, moderate, low, very low)⁸³. For validated tools (e.g., QuADS, MI-CLAIM), we calculated the percentage of maximum score (High ( ≥ 80%), Moderate (60–79%), Low (40–59%), and Very low ( < 40%)). For AACODS, we graded quality by the number of domains met (High 5–6, Moderate 3–4, Low 1–2, Very low 0).

Author contributions

M.T.Y., L.N.Z., and L.L.Z. conceived and designed the scoping review. LNZ and LLZ supervised the study. M.T.Y., L.N.Z., Y.L., and T.H. performed literature search, screening, and data analysis. M.T.Y. and L.N.Z. drafted the first version of the manuscript. L.N.Z., L.L.Z., S.D., H.L.L., K.Z., G.S.H., P.O.V., L.K.D., D.A.D., K.D., X.X.Z., and W.Z. critically reviewed and revised the manuscript. All authors had full access to the data, interpreted the findings, and approved the final manuscript.

Data availability

All data generated or analyzed during this study are included in this published article and its Supplementary Information files. The full extracted dataset underlying the analyses (e.g., study characteristics and AI classification results) is provided in the Supplementary Tables to enable replication and further exploration.

Code availability

This study did not involve the utilization of any custom code or mathematical algorithm.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41746-025-02156-2.