Utilization of artificial intelligence in clinical practice: A systematic review of China's experiences

Abstract

Background

Artificial intelligence (AI) is transforming clinical applications, including diagnostics, treatment planning, drug discovery, and administrative tasks. Despite significant progress, AI remains a double-edged sword, and its implementation requires careful, evidence-based evaluation. To date, few AI applications have been fully integrated into clinical workflows, especially in significant populations.

Objective

This study aims to synthesize evidence on AI utilization in clinical practice, identify key facilitators and barriers, and provide recommendations for implementation within relevant sociocultural and demographic contexts.

Methods

Following PRISMA guidelines, this review conducted a comprehensive search in Web of Science, Scopus, and PubMed. Bias was assessed using the JBI and NOS tools. Data on study design, population, AI technologies, applications, clinical issues, and outcomes were extracted. Emerging themes were organized using the NASSS framework.

Results

Of 1002 records screened, 28 studies were included, most of which were cross-sectional (57%). Machine learning (ML) (43%) was the most frequently used AI technology. AI application outcomes primarily focused on application performance (61%), clinical outcomes (43%), and patient outcomes (32%). Clinical contexts included infectious diseases, chronic conditions, imaging, and physician–patient interactions. Key facilitators included perceptions of operational efficiency, availability of AI tools, confidence in improved accuracy, alignment with goals, perceived cost-saving potential, and enabling environments. Reported barriers involved ethical and privacy concerns, limited user acceptance, inconsistent accuracy, technical complexity, unclear accountability, trust-related issues, and inadequate infrastructure.

Conclusions

AI in clinical practice holds tremendous potential in diagnostic accuracy, workflow efficiency, patient engagement, and cost-effectiveness. AI-assisted approaches perform at least as well as conventional methods, even better. Key characteristics within specific contextual settings were synthesized, and contextually informed recommendations were proposed to facilitate AI integration and address the identified barriers. Future research should focus on evaluating AI's long-term impact and addressing emerging issues as AI becomes more embedded in clinical workflows.

Introduction

Background

Artificial intelligence (AI), a transformative technology that enables computers to perform a variety of advanced functions and the backbone of modern computing innovation that can deliver value. ¹ It encompasses various subfields, including machine learning (ML), deep learning (DL), natural language processing (NLP), and computer vision (CV). Now, AI technology has been applied across various domains, including healthcare for disease diagnosis, ² virtual assistants such as AI-powered Alexa, ³ finance for fraud detection, ⁴ and transportation with self-driving cars, ⁵ to name a few.

Among the fields where AI is being applied, clinical practice stands out due to its potential to address critical challenges. Clinical or health issue is a complex, iterative process that requires professionals to continuously collect, analyze, and synthesize data from diverse sources while applying clinical expertise and evidence-based research to inform decision-making and enhance patient care. ⁶ Despite these efforts, clinical practice faces significant challenges, such as medical errors resulting from overwhelming workloads and incomplete information,^7–9 shortages of medical staff that impact the quality of care, and unequal distribution of medical resources in remote or underserved areas. ¹⁰ Rising treatment costs further hinder access to care,^11,12 and the early diagnosis of diseases such as cancer and cardiovascular conditions remains challenging due to vague or nonspecific symptoms. ¹³ The integration of AI into clinical issue is poised to drive transformative changes, with projections showing a compound annual growth rate of 31% in the AI-driven clinical sector from 2022 to 2027. ¹⁴ AI has already demonstrated significant progress across various aspects of clinical issues, including diagnostics, ¹⁵ treatment planning, ¹⁶ drug discovery, ¹⁷ and administrative efficiency, ¹⁸ among others. Despite rapid advancement of AI, the integration of it into healthcare remains a double-edged sword. On one hand, AI technologies offer opportunities in technological advancements, decision making, and diagnosis and patient monitoring. ¹⁹ On the other hand, challenges persist regarding ethical and privacy, lack of awareness, unreliability and trustworthiness, and healthcare providers liability. ¹⁹ These highlight the need for an evidence-informed assessment of AI implementation in clinical environments, particularly within complex health systems.

Research gaps and objectives

Numerous reviews have explored AI in clinical field from different perspectives, such as disease diagnosis, ²⁰ image mining, ²¹ pregnancy care, ²² and dentistry. ²³ Beyond clinical perspectives, other reviews have focused on broader dimensions of AI in healthcare, such as its effectiveness, ²⁴ economic impacts, ²⁵ and implications for business and management. ²⁶ Despite the growing literature on AI in clinical practice, research gaps remain. Most existing reviews tend to focus on narrowly defined clinical or operational aspects, often limiting the scope of their insights to specific applications or perspectives, which may restrict a broader understanding of AI's potential across various clinical issues. Furthermore, there is a lack of comprehensive evaluations that encompass a wide range of outcomes, facilitators and barriers associated with AI implementation, particularly those addressing the sociocultural and demographic factors unique to large, globally significant populations.

In the development of AI in clinical fields, China has made its position as a key player in the global landscape, employing technologies such as intelligent robotics and virtual reality across various sectors, including clinical practice. Evidence suggests that China ranks second globally in the number of research publications on AI in healthcare ²⁷ and has extensively integrated AI into clinical applications, with substantial research conducted in this domain. ²⁸ The progress is made by strategic government policies and substantial investments, with the nation aiming to become a global AI innovation hub by 2030. By that time, China's core AI industry is projected to exceed 1 trillion yuan (approximately $140.9 billion). ²⁹ Notably, as the world's largest developing country, China faces healthcare challenges such as urban-rural resource disparities, an aging population, and a rising burden of chronic diseases, which closely parallel those encountered in other developing nations. Consequently, China's AI-integrated clinical practices may serve as replicable models for resource-constrained settings. Furthermore, its experience holds particular relevance for global ethnic Chinese communities, which often share cultural values, healthcare-seeking behaviors, and socioeconomic patterns that align with China's health context.

This review aims to synthesize current research on the utilization of AI in clinical practice, focusing on China's experiences. Specifically, the review seeks to achieve the following objectives: first, to integrate AI utilization in clinical settings with existing literature and identify key clinical issues and outcomes addressed through various AI technologies; and second, to assess the key facilitators and barriers of AI utilization under a comprehensive framework. Considering the gap between research and application, this review also proposes contextually informed recommendations to enhance AI integration and focuses on its implementation in clinical practice, with attention to the sociocultural and demographic contexts of ethnic Chinese communities worldwide and other developing regions with characteristics comparable to those of China.

Methods

Study design

This review was conducted in accordance with PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines ³⁰ (Multimedia Appendix 1), and the protocol of this systematic review was not registered. This review addressed two research questions: (a) What are the types of AI utilization in clinical practice, including study characteristics, clinical issues addressed, and evaluation outcomes? and (b) What are the key facilitators and barriers of utilizing AI in clinical practice?

Eligibility criteria

Studies published in English between January 2010 and September 2024 that focused on the utilization of AI in clinical practice in China were considered for inclusion. English-language studies were selected because English was the predominant language for global scientific communication, and it facilitated access to a wide range of peer-reviewed research. Additionally, AI research in healthcare began gaining significant momentum around 2010. In this review, the term “AI” refers to a spectrum of technologies, including ML, DL, NLP, and CV, all of which are increasingly integrated into clinical practice to enhance diagnostic accuracy, decision support, and patient management. This review included studies that if they (a) were original research focusing on the utilization of AI technologies in clinical practice, either involving patients, healthcare providers, or healthcare settings, (b) reported findings on feasibility and usability, and (c) were conducted within the context of China and published in English. In meanwhile, studies were excluded if they (a) only focused on the design of clinical AI algorithms without practical implementation, (b) only focused on medical hardware devices, such as ultrasound equipment, etc., and (c) were reviews, conference abstracts, commentaries, and ongoing studies.

Information source

A comprehensive search was performed across the Web of Science, Scopus, and PubMed databases to identify studies published from January 2010 to September 2024.

Search strategy

The search strategy of this review employed a PICO-based search string, incorporating keywords relevant to AI, clinical practice, and China. Terms included “artificial intelligence,” “AI,” “artificial general intelligence,” “machine intelligence,” “machine learning,” “deep learning,” “clinical,” “healthcare,” “health,” “medical care,” “implement,” “implementation,” “China,” “Chinese,” etc. The search terms were detailed in Multimedia Appendix 2. Two researchers (YQ and CZ) independently conducted the screening and identification of studies. Discrepancies were addressed through discussion, and, if necessary, consensus was reached with input from other researchers (EM and AAA).

Selection process

Two researchers (YQ and CZ) independently imported all retrieved studies into EndNote, where duplicate records were identified and removed. They then independently screened the titles and abstracts of the remaining studies to assess initial eligibility. To ensure consistency, any initial disagreements such as differing judgments on whether a study met the inclusion criteria, were discussed between the two researchers. If consensus could not be reached at this stage, the case was referred to a third reviewer (EM), who independently assessed the disputed study and made a final decision. Full-text articles of potentially eligible studies were then independently reviewed by the same two researchers (YQ and CZ). A predefined approach was employed to resolve any disagreements during the full-text screening. These disagreements were typically related to the interpretation of eligibility criteria, such as whether a study should be classified as focusing on the design of clinical AI algorithms without practical implementation. In such cases, the two researchers (YQ and CZ) first attempted to reach consensus through discussion. If disagreement remained, the issue was moved to a consensus meeting involving two additional researchers (EM and AAA), who independently assessed the study. Final decisions were made through group discussion to ensure consistency and minimize selection bias.

Data collection, data items, and data synthesis

To ensure consistency, two researchers (YQ and CZ) independently piloted the data extraction process using a standardized, pre-tested extraction form applied to a sample of three studies. They compared their results, resolved discrepancies through discussion, and subsequently refined the extraction framework. After this process, YQ extracted data from the remaining studies, while CZ independently reviewed a randomly selected subset to assess reliability. All extracted data were collaboratively reviewed, with disagreements resolved through consensus or adjudicated by a third researcher (EM).

For RQ1, the extraction form captured both descriptive study characteristics and implementation-related outcomes. Extracted variables included citation details (author, title, and publication date), study design, and population characteristics. Information on AI utilization was also collected, such as the type of AI technology, its specific application, and the targeted clinical issues. In accordance with the clinical evaluation outcomes framework proposed by Yin et al., ²⁸ evaluated outcomes were synthesized and categorized into four domains, that were, application performance (AP), clinician outcomes (CO), patient outcomes (PO), and cost-effectiveness (CE).

For RQ2, this review adopted a hybrid approach to identify the facilitators and barriers associated with the utilization of AI in clinical practice across included studies. Initially, an inductive strategy was employed to identify emerging themes and systematically document the corresponding records. This process was subsequently guided by the Non-adoption, Abandonment, Scale-up, Spread, and Sustainability (NASSS) framework ³¹ (Figure 1), which offered a structured lens for evaluating the implementation of digital health technologies. A deductive process then followed to ensure that all 7 NASSS domains were comprehensively addressed: the condition, the technology, the value proposition, the adopter system, the organization, the wider system, and embedding and adoption over time. This hybrid approach has been utilized and verified in previous implementation research. ³² A narrative synthesis was then conducted to summarize these findings and contextualize them within each NASSS domain.

Figure 1.

The NASSS (Non-adoption, Abandonments, Scale-up, Spread, and Sustainability) framework.

Due to substantial heterogeneity in study designs, outcomes, and settings, a meta-analysis was not feasible. Instead, a narrative synthesis approach was employed, supported by thematic analysis within each framework domain. All data extraction and synthesis activities were managed using Microsoft Excel 2023 to ensure transparency and completeness. A conceptual framework map was presented in Figure 2.

Figure 2.

Conceptual framework map. RQ: research question. NASSSS: Non-adoption, Abandonment, Scale-up, Spread, and Sustainability.

Risk of bias assessment and reporting bias assessment

Two researchers (YQ and CZ) independently assessed risk of bias in the included studies, with findings cross-verified. Any discrepancies were resolved through consultation with other researchers (EM and AAA). Given the heterogeneity in study designs, we employed appropriate risk of bias assessment tools for most included studies. The Joanna Briggs Institute (JBI) Critical Appraisal Checklist was used for cross-sectional studies, while the Newcastle–Ottawa Scale (NOS) was applied to retrospective cohort studies. The JBI checklist consists of 8 domains, evaluating whether a study clearly defines inclusion criteria, details the study population and setting, employs valid and reliable exposure and outcome measurements, applies objective and standardized assessment criteria, identifies and addresses confounding factors, and utilizes appropriate statistical analyses. Each domain was rated as “yes,” “no,” or “unclear.” Studies with a total score of 7 or higher were classified as low risk, those scoring between 4 and 6 as moderate risk, and those scoring 3 or lower as high risk of bias. The NOS tool evaluates selection bias (4 points), comparability bias (2 points), and outcome bias (3 points), with a total score ranging from 0 to 9. A higher score indicates a lower risk of bias. Studies with a total score of 7 or higher were classified as low risk, those scoring between 5 and 6 as moderate risk, and those scoring below 5 as high risk of bias.

Results

Study screening

The initial search identified 1002 studies in English from PubMed, Web of Science, and Scopus. After removing 595 duplicates, 407 studies proceeded to the screening stage. After reviewing the title and abstract of these studies, 293 studies were excluded because of they were not related the utilization of AI technologies in clinical practice in China, resulting 114 studies were selected for full-text screening. Of these, 86 studies were excluded due to only focused on the design of clinical AI algorithms without practical implementation, only focused on medical hardware devices, and was a review, conference abstracts, commentaries, and ongoing studies. Ultimately, 28 studies were included in the final synthesis. A PRISMA flowchart of this process was detailed in Figure 3.

Figure 3.

PRISMA (Preferred Reporting Items for Systematic reviews and Meta-Analyses) flowchart.

Study characteristics

The literature search identified 28 eligible studies, the characteristics of which were summarized in Table 1, while a summary of them was presented in Table 2. These studies encompassed various research designs, including cross-sectional studies (16/28, 57%), retrospective cohort studies (8/28, 29%), qualitative studies (2/28, 7%), randomized controlled trials (1/28, 4%), and non-randomized controlled trials (1/28, 4%). Notably, more than half of these studies (15/28, 54%) were conducted in 2024. In terms of AI technology utilization, ML was the most commonly employed approach (12/28, 43%), followed by NLP (7/28, 25%), DL (5/28, 18%), and CV (4/28, 14%). Regarding evaluation outcomes, the majority of studies (17/28, 61%) assessed AP, while 12 (43%) focused on CO, 9 (32%) examined PO, and 6 (21%) investigated CE. Additionally, an analysis of study populations revealed that 9 studies (32%) included medical professionals, such as nurses, specialists, oncologists, physicians, health information technicians, radiology residents, and healthcare workers, whereas 8 studies (29%) explicitly reported patient samples. The clinical issues addressed in the reviewed studies showed a diverse range of medical conditions, diagnostic approaches, and healthcare processes, including infectious diseases, chronic illnesses, medical imaging, diagnostic triage, physician and patient interactions, and administrative tasks.

Table 1.

Characteristics of the eligible studies.

Author and years	Study design (risk of bias)	Sample characteristics	AI technology utilized	Clinical issues addressed	Evaluation of key outcomes
K. Zhang et al. 2020 33	Retrospective cohort study (low)	3777 patients	ML (a developed AI system)	Novel coronavirus pneumonia	AP (accuracy) CO (rapid diagnosis) CE (reduce patient waiting time, clinician's workload)
X. Li et al. 2021 34	Retrospective cohort study (low)	12342 patients	DL (a developed AI-assisted module named XIAO YI)	Imaging examinations or laboratory tests	PO (reduce patient waiting time) CE (enhances hospital outpatient services)
Sun 2021 35	Qualitative study (not applicable)	29 interviews from 3 hospitals	DL (4 medical AI systems provided by iFlytek)	Administrative tasks, diagnosis and imaging	CE (the expert and boss strategies suit AI systems)
Zeng et al. 2021 36	Randomized controlled trial (not applicable)	150 patients	ML (a personalized AI system for sports management)	Chronic diseases, and diabetes	PO (optimize care-related activities and health management)
Cheng et al. 2022 37	Cross-sectional study (moderate)	343 healthcare workers	ML (a service model of AI-assisted diagnosis and treatment)	Dental diseases	AP (accuracy) CO (rapid diagnosis)
Huang et al. 2022 38	Non-randomized controlled trial (not applicable)	32 people and 6 nurses	CV (an AI-based surveillance system)	Nurse–patient interaction	CO (intuitively assess patients’ status, reducing effort in detecting emergencies) PO (avoids secondary injuries) CE (reducing the interaction time)
Leng et al. 2022 39	Cross-sectional study (moderate)	Real patients’ records	ML (a Bi-level AI model)	Respiratory diseases	AP (accuracy)
Q. Liu et al. 2022 40	Retrospective cohort study (low)	127031 participants	ML (an AI prediction model)	Type 2 diabetes mellitus	AP (accuracy, predictive value)
X. Liu et al. 2022 41	Cross-sectional study (low)	543 patients	DL (an AI-powered service robot)	Self-service agents	PO (enhance trust, ease-of-use, and enjoyment)
Chen et al. 2023 42	Cross-sectional study (low)	3666 radiology residents	ML (AI system)	Radiology residents’ perceptions	AP (accuracy) CO (reduce the demand)
Wong et al. 2023 43	Qualitative study (not applicable)	6 physicians and 24 patients	DL (a digital detection surveillance)	Infectious diseases	CO (decision making) PO (informed consent, discrimination or social inequity against) CE (maintaining data security)
Y. Wang et al. 2023 44	Cross-sectional study (low)	137 pediatricians, 135 nurses and 60 health information technicians	ML (a medical AI)	Pediatric healthcare	CO and PO (knowledge gaps and attitude differences existed)
X. Zhu et al. 2023 45	Retrospective cohort study (low)	1,518,972 women	CV (an online cytological assessment system using AI)	Cervical cancer	CE (low-cost, accessible and effective)
Cao et al. 2024 46	Cross-sectional study (low)	677 medical staffs	ML (an AI triage system)	Medical triage	CO (rapid diagnosis)
M. Guo et al. 2024 47	Cross-sectional study (moderate)	400 physicians	ML (an augmentation and automation AI model)	Physicians’ performance	CO (augmentation AI use increased productivity and innovation, while automation AI use had the opposite effect)
M. Li et al. 2024 48	Cross-sectional study (low)	228 oncologists	NLP (AI system)	Oncologists’ perspectives	AP (transparency, human-centered design, bias mitigation, and AI education)
W. Li and Wang 2024 49	Cross-sectional study (low)	604 users	CV (an AI-assisted diagnostic system)	Users' attitudes and reasons	CO and PO (increase AI-assisted diagnostic system adoption)
Liang et al. 2024 50	Retrospective cohort study (moderate)	4386 subjects	ML (a machine learning model)	Thyroid nodules	AP (accuracy, predictive value)
F. Lin et al. 2024 51	Retrospective cohort study (high)	22714 patients	ML (an AI-based mechanism, called HBBI-AI)	Atrial fibrillation	AP (predictive value)
S. Lin et al. 2024 52	Cross-sectional study (low)	358 residents	CV (a facilitated self-service eye screening pattern)	Eye disease	AP (accuracy)
Q. Liu et al. 2024 53	Retrospective cohort study (moderate)	Dataset, n=18927	ML (2 AI assessment models)	Bone age	AP (accuracy)
Tao et al. 2024 54	Retrospective cohort study (low)	28728 individuals	DL (an AI diagnostic system)	Lung cancer	AP (high participation and early detection rates)
Wang et al. 2024 55	Cross-sectional study (low)	350 hypertensives	NLP (AI telephone calls)	Hypertension	AP (efficiency, data quality) CO (disease management, reduce workload and time)
T. Wang, Mu et al. 2024 56	Cross-sectional study (low)	272 clinical nurses	NLP (ChatGPT)	Tracheostomy care	AP (accuracy) CO and PO (knowledge improvement)
Y. Wang, Chen et al. 2024 57	Cross-sectional study (low)	96 inquiries	NLP (ChatGPT)	Chronic hepatitis B	AP (accuracy, disease management)
Yang et al. 2024 58	Cross-sectional study (moderate)	590 hospitals	NLP (intelligent guidance chatbots)	Medical guidance and triage	AP (usability) CO (optimizing appointment services, alleviating pressure in triage) PO (enhancing experience)
Ye et al. 2024 59	Cross-sectional study (low)	431 participants	NLP (ChatGPT)	Allergic rhinitis and chronic rhinosinusitis	AP (accuracy)
Q. Zhang et al. 2024 60	Cross-sectional study (low)	33 questions with 3 specialists	NLP (ChatGPT)	Gingival and endodontic health	AP (accuracy, comprehensiveness)

AP: application performance; CO: clinician outcomes; PO: patient outcomes; CE: cost-effectiveness; ML: machine learning; DL: deep learning; NLP: natural language processing; CV: computer vision.

Table 2.

Summary of the eligible studies.

Items	Proportion, n (%)
Research type
Cross-sectional study	16 (57)
Retrospective cohort study	8 (29)
Qualitative study	2 (7)
Randomized controlled trial	1 (4)
Non-randomized controlled trial	1 (4)
Publication year
2024	15 (54)
2022	5 (18)
2023	4 (14)
2021	3 (11)
2020	1 (4)
AI technology utilized
ML	12 (43)
NPL	7 (25)
DL	5 (18)
CV	4 (14)
Evaluation key outcomes
AP	17 (61)
CO	12 (43)
PO	9 (32)
CE	6 (21)

AP: application performance; CO: clinician outcomes; PO: patient outcomes; CE: cost-effectiveness; ML: machine learning; DL: deep learning; NLP: natural language processing; CV: computer vision.

Bias reporting

Risk of bias was assessed using the JBI checklist for 16 studies and the NOS for 8 studies. Detailed assessment results for assessed studies were provided in Multimedia Appendix 3. Of the 16 studies evaluated using the JBI checklist, 12^{41,42,44,46,48,49,52,55–57,59,60} were classified as low risk of bias, and 4^37,39,47,58 were found to have a moderate risk of bias. Although most studies effectively applied objective and standardized assessment criteria, 1 ⁴⁴ did not define criteria for AI perception and acceptance beyond survey items. Four^37,39,47,58 failed to explicitly identify potential confounders or describe strategies for addressing them. Additionally, 1 ³⁹ did not provide sufficient reliability testing for outcome measures. Among the 8 studies using the NOS, 5^{33,34,40,45,54} were classified as low risk of bias, 2^50,53 exhibited moderate risk of bias, and 1 ⁵¹ study had a high risk of bias. The primary sources of bias were insufficient adjustment for confounding variables in comparability assessments^50,53 and outcome measurement issues due to reliance on retrospective electronic medical records (EMR) ⁵⁰ or AI-based automated classification. ⁵¹ Overall, 17 of the 24 studies assessed (71%) were classified as having a low risk of bias, suggesting a generally methodological rigor. Relevant results were presented in Table 1.

AI technologies utilized in addressing clinical issues

Across the 28 eligible studies on AI utilization in clinical practice, AI technologies were classified into four primary categories: ML (12/28, 43%), NLP (7/28, 25%), DL (5/28, 18%), and CV (4/28, 14%). They demonstrated applications across clinical issues, such as diagnostic accuracy, clinical decision support, patient management, and healthcare workflow optimization.

ML was the most frequently used (12/28, 43%), primarily applied in predictive modeling, disease diagnosis, and decision-making. It supported diagnostic tasks in conditions such as thyroid nodules, ⁵⁰ atrial fibrillation, ⁵¹ type 2 diabetes mellitus, ⁴⁰ and bone age assessment. ⁵³ ML-based tools also facilitated personalized chronic disease management, ³⁶ improved diagnosis and treatment in dentistry, ³⁷ enhanced pediatric care, ⁴⁴ and optimized triage workflows. ⁴⁶ Moreover, ML was used in studies evaluating its impact on physician performance, ⁴⁷ radiology residents’ AI perceptions, ⁴² and respiratory disease diagnostics. ³⁹ NLP was utilized in 7 of 28 studies (25%), mainly for triage, AI-driven consultation, and patient communication. It streamlined hypertension management via telephone consultations, ⁵⁵ enabled patient education through ChatGPT in tracheostomy care, ⁵⁶ and was applied in managing chronic hepatitis B, ⁵⁷ and rhinitis. ⁵⁹ NLP also supported workflow in hospitals, ⁵⁸ evaluated AI transparency in oncology, ⁴⁸ and were used in dental health assessment. ⁶⁰ DL, applied in 5 of 28 studies (18%), played roles in medical imaging and diagnostic support. Examples include AI-assisted radiological imaging, ³⁴ self-service healthcare robots, ⁴¹ digital infectious disease surveillance, ⁴³ lung cancer screening, ⁵⁴ and AI systems combining administrative and diagnostic functions. ³⁵ CV appeared in 4 of 28 studies (14%) and was integrated into diagnostic, monitoring, and screening applications. Use cases included AI-supported cervical cancer screening, ⁴⁵ eye screening, ⁵² nurse–patient interaction enhancement, ³⁸ and diagnostic adoption facilitation. ⁴⁹

Temporal trends in AI technologies utilization were evident across the included 28 studies. As shown in Table 2, more than half (15/28, 54%) were published in 2024, followed by 5 studies in 2022, 4 in 2023, 3 in 2021, and 1 in 2020. Among the 9 studies published between 2020 and 2022, the majority focused on the development and implementation of ML (5/9, 56%), followed by DL (3/9, 33%). These AI tools were typically used for diagnostic support, predictive modeling, and clinical workflow optimization. ML accounted for the majority of implementations during this period, with applications such as bi-level AI models, ³⁹ personalized sports management, ³⁶ and AI-assisted diagnosis. ³⁷ DL-based tools, including hospital-developed modules like AI-assisted module named “XIAO YI” ³⁴ and AI-powered service robots, ⁴¹ also supported clinical diagnostics and automation. Among 9 studies, 1 (11%) study explored CV for surveillance-based applications. ³⁸ In comparison, 19 studies were published between 2023 and 2024, showing a shift toward more diverse and user-centered AI technologies. Notably, NLP (7/19, 37%) emerged as commonly applied, supporting applications such as chatbot consultations,^56–60 AI-driven telephone follow-ups, ⁵⁵ and intelligent guidance systems. ⁴⁸ ML (7/19, 37%) also remained widely used, though its focused transitioned from model development to supporting triage systems, ⁴⁶ disease management, ⁴⁴ and system integration. ⁴² CV was also increasingly adopted (3/19, 16%) to enhance service diagnostic tools^45,49 and screening programs. ⁵² Meanwhile, DL applications (2/19, 11%) were integrated into digital surveillance ⁴³ and diagnostic decision-making, ⁵⁴ often in combination with other technologies.

Key outcomes evaluated

This review identified key outcomes categorized into AP (17/28, 61%), CO (12/28, 43%), PO (9/28, 32%), CE (6/28, 21%). These outcome dimensions were systematically analyzed in conjunction with the specific AI technologies used in the 28 included studies.

A total of 17 studies^{33,37,39,40,42,48,50–60} evaluated the outcomes of AP in AI-assisted clinical practice. Among them, 11 studies reported that AI utilization demonstrated high diagnostic accuracy. For instance, Zhang et al. ³³ found that AI-powered solution employing ML algorithms could significantly enhance clinical prognosis accuracy. Their study demonstrated that the AI system differentiated novel coronavirus pneumonia from other types of pneumonia and normal controls with an accuracy of 92.49%, a sensitivity of 94.93%, and a specificity of 91.13%. The overall performance achieved 92.49% accuracy with an area under the receiver operating characteristic of 0.9813, indicating robust diagnostic capability. According to Liu et al., ⁵³ ML-based AI assessment models demonstrated high accuracy in predicting bone age for both Han and Tibetan children. Their findings indicated that the model achieved an accuracy within 1 year of 97.7% on the local test set and 89.5% on the external test set. In contrast, Wang et al. ⁵⁶ reported that while the NLP-based ChatGPT-4.0 made higher accuracy than ChatGPT-3.5, it still had certain limitations. For instance, in the domain of tracheostomy and peristomal skin care, its accuracy was lower than that of clinical nurses participating in the study. Nonetheless, the study highlighted that ChatGPT could serve as a supplementary tool for medical information. In addition to demonstrating high diagnostic accuracy, AI performance was also evaluated in terms of predictive value, usability, efficiency, comprehensiveness, and other metrics. Lin et al. ⁵¹ demonstrated that ML-based atrial fibrillation risk prediction, which relied on heartbeat interval data, was successfully integrated with wearable devices to enable real-time monitoring and early detection. Yang et al. ⁵⁸ investigated the usability of the NLP-based hospital intelligent guidance chatbots, identifying key influencing factors such as functional diversity, human-like characteristics, outpatient volume, and staff size, which impacted adoption and user satisfaction. Additionally, Wang et al. ⁵⁵ highlighted the feasibility of NLP-based AI-powered follow-up technology in community-based chronic disease management, reporting that an AI-driven follow-up platform significantly enhanced efficiency by assisting healthcare providers in plan formulation, phone consultations, information collection, structured data storage, and data uploading.

Twelve studies^{33,37,38,42–44,46,47,49,55,56,58} investigated the outcomes of CO in AI-assisted clinical practice, and addressed a diverse range of aspects. Three studies reported that AI can facilitate rapid diagnosis in clinical settings. Specifically, Zhang et al. ³³ demonstrated that an AI system employing ML algorithms assists radiologists and physicians in making prompt diagnoses, especially in overloaded healthcare systems. Similarly, Cheng et al. and Cao et al.^37,46 found that diagnosis and treatment supported by ML-driven AI could broaden clinical applications and accelerate workflow. Three studies^44,49,56 highlighted the importance of knowledge regarding AI technologies, including ML, CV, and NLP. Four studies^38,42,55,58 indicated that AI-assisted clinical practice contributed to reducing clinicians’ workload across multiple domains. Huang et al. ³⁸ demonstrated that CV-based AI reduced the effort required to detect emergencies, thereby streamlining urgent care processes. Similarly, Chen et al. ⁴² found that integrating ML-based AI into clinical workflows diminished the demand on healthcare resources. In addition, Wang et al. ⁵⁵ reported that NLP-based AI reduced both the workload and the time required for clinical tasks, while Yang et al. ⁵⁸ observed that it also alleviated pressure during patient triage. Notably, Guo et al. ⁴⁷ reported that augmentation AI systems powered by ML enhanced clinicians’ productivity and innovation, whereas automation-driven AI systems were linked to adverse outcomes.

Nine studies^{34,36,38,41,43,44,49,56,58} reported the outcomes of PO in AI-assisted clinical practice. Among these, four studies emphasized the optimization and efficiency of patient-related outcomes. Specifically, Zeng et al. ³⁶ focused on the ML optimization of care-related activities and health management, Li et al. ³⁴ highlighted the role of DL in reducing patient waiting times, Li and Wang ⁴⁹ examined the increased adoption of CV-based AI-assisted diagnostic systems. Wang et al. ⁵⁷ reported that the NLP-based AI system improved the overall patient experience. In addition to these four studies, several others explored various dimensions of PO. For instance, Liu et al. ⁴¹ demonstrated that DL-based AI systems enhanced users’ trust, ease of use, and enjoyment, thereby improving user engagement and satisfaction. Wang et al. ⁴⁴ and Wang et al. ⁵⁶ identified knowledge gaps and attitude differences among patients, especially those involving ML and NLP technologies. According to Wang et al., ⁴⁴ only 2.4% to 9.6% of participants were familiar with the ethical implementation of ML-based medical AI in social experiments, while 31.9% to 86.1% of participants expressed agreement with ethical implementation attitudes. In contrast, Wang et al. ⁵⁶ highlighted the potential of NLP-based AI as a complementary medical information tool for patients, and argued its effectiveness in improving patients’ understanding of tracheotomy care.

There were 6 studies^{33–35,38,43,45} reported the outcomes of CE in AI-assisted clinical practice. Notably, 3 of these studies demonstrated that AI can reduce costs by improving time efficiency. Specifically, AI systems utilizing ML, DL, and CV have been shown to decrease patient waiting times and clinicians’ workloads, ³³ enhance the efficiency of hospital outpatient services, ³⁴ and shorten interaction times, ³⁸ which lead to significant cost savings. In addition, other studies have highlighted additional factors influencing CE in AI-assisted clinical practice. For example, Zhu et al. ⁴⁵ demonstrated that low-cost, accessible, and effective CV-based AI solutions was able to improve economic outcomes by broadening their reach and reducing expenditures. Wong et al. ⁴³ emphasized the role of DL in maintaining data security to prevent costly breaches, thereby safeguarding both financial and clinical resources. Furthermore, Sun ³⁵ found that implementing expert and boss strategies within DL-based AI systems was able to optimize decision-making processes, enhance operational efficiency, and contribute to a more cost-effective clinical workflow.

Reported facilitators and barriers

Across the 28 included studies, a total of 56 records of potential facilitators and 48 records of potential barriers to utilizing AI in clinical practice were identified and documented. Facilitators and barriers were reported in all included studies. The records were synthesized and categorized according to the 7 domains of the NASSS framework ³¹ (Figure 1).

Domain 1: The condition

Domain of the condition addresses both the clinical and sociocultural dimensions of the health condition and its comorbidities, recognizing that the effectiveness of health technologies may vary among individuals with the same diagnosis. ³¹ Facilitators for AI applications were commonly identified in clinical practice involving structured, high-burden, or standardized diseases. This theme was reported in the included studies, which commonly referenced conditions such as COVID-19 pneumonia, thyroid nodules, diabetic complications, and hypertensive disorders.^{33,36,39,45,51,55} These diseases often presented with defined features or quantifiable indicators, enabling AI systems to support early detection, triage, or risk stratification. Chronic diseases with stable progression, such as type 2 diabetes, chronic hepatitis B, chronic rhinosinusitis, and childhood growth disorders, were also consistent with AI-based tools due to their suitability for predictive modeling and routine follow-up.^40,53,57,59 In particular, clinical areas involving children and dentistry benefited from task standardization, which supported consistency in AI deployment.^37,44,60 Furthermore, evidence suggested that AI tools added value to screening initiatives targeting common diseases like cervical and lung cancer, where scalable diagnostic solutions were needed.^45,54

Greater condition complexity could reduce the clinical appropriateness of AI applications and restrict their effectiveness. ³¹ Such complexity, as well as diagnostic subjectivity, frequently contributed to implementation barriers. AI struggled to adapt to multifactorial conditions involving complications, such as cancer, skeletal disorders, or viral hepatitis with etiologies, where expert judgment remained indispensable.^33,48,53,57 Additionally, conditions that required symptom interpretation, such as asthma exacerbation, unexplained abdominal pain, or psychiatric symptoms, often exceeded the interpretative capacity of AI systems, which limited their perceived clinical value.^34,35,56 Beyond clinical complexity, patient engagement and disease awareness also played a critical role in determining the suitability of AI applications. In cases where users had poor understanding of their condition, incomplete data inputs diminished model performance and intervention efficacy.^36,55 Similarly, the reliability of AI-based alerts that relied on self-reported symptoms for infectious disease surveillance was compromised, as underreporting occurred due to fear of isolation. ⁴³

Domain 2: The technology

The technology domain covers aspects such as materials, data, knowledge, and supply-related characteristics. It considers both the physical tools and user expertise involved, as well as the mutual influence between the technology and its context of use. ³¹ In AI applications, facilitators included high diagnostic accuracy, real-time processing speed, and compatibility with existing data formats and clinical systems. Evidence showed that AI models demonstrated strong technical performance, particularly in image-based diagnostics, through the use of training datasets, DL architectures, and integration with standard imaging modalities.^33,34,48 Building on this foundation, modular systems incorporating wearable devices, or voice-enabled EMRs further supported data acquisition and real-time feedback, thereby improving efficiency in clinical settings.^36,55

Ease of use and accessibility were also facilitators. Interfaces embedded into familiar platforms (e.g. WeChat mini-programs) supported interaction in clinical settings.^34,41,58 Studies also highlighted that interactive and visually guided designs enhanced users’ perceptions of usefulness and ease of use, which in turn positively influenced acceptance among both clinicians and patients.^41,58

However, significant technical barriers persisted. A major limitation was called “black box” nature of many AI algorithms, which lacked transparency, thereby making clinician skepticism and reducing trust.^33,37,48 In such systems, the decision-making process was often opaque, making it difficult for users to understand how inputs were translated into recommendations. This lack of transparency not only limited accountability but also raised concerns regarding diagnostic accuracy, and the ability to validate results in complex or high-risk cases. Although techniques such as the Shapley additive explanations ⁴⁰ and Python ⁵⁰ were employed to improve interpretability, many complex models remained difficult for clinicians to interpret. The problem was especially critical in high-risk fields such as oncology, where clear clinical reasoning and accountability are essential. ⁴⁸ Integration challenges also limited adoption. Systems not fully embedded into hospital information infrastructures required redundant data entry or additional effort from clinicians, which might reduce usability.^35,37 In some cases, AI tools required specific hardware, stable internet connectivity, or experience imaging devices, but these conditions were not available across rural or low-resource settings.^36,45,54

Domain 3: The value proposition

The value proposition domain assesses whether a technology offers sufficient perceived benefit to justify its development and adoption by clinicians, patients, and suppliers. ³¹ It affects both the supply-side investment and demand-side acceptance. In clinical AI applications, perceived value was most commonly linked to improvements in diagnostic accuracy, workflow efficiency, and healthcare accessibility. Studies reported that AI tools enhanced clinical decision-making by reducing diagnostic errors, shortening waiting times, and reducing provider burden in primary care and high-demand settings.^33–35,46 In low-resource environments, AI systems offered potential for expanding care coverage among underserved populations.^36,45,53,54

Cost-effectiveness was also a contributor to the perceived value of AI in clinical settings. Several studies reported that AI implementation reduced direct healthcare expenditures, including testing fees, registration costs, and staffing needs in automation and triage support.^34,39,52,55 In addition, AI tools for patient like chatbots and digital guidance systems were valued for improving patient experience and enhancing access to information.^41,56,60 These features, as reported in the studies, were found to be beneficial in pediatric care, chronic disease management, and high-volume hospital environments.

Despite the technical potential of AI, barriers to establishing a strong perceived value proposition persisted. These barriers reflected both functional limitations and broader contextual concerns that influenced the willingness to adopt AI tools. A key limitation was that many AI tools were designed for narrow tasks and were not suitable for comprehensive clinical decision-making. This constrained their perceived utility among clinicians, who continued to rely on human judgment for decisions.^34,37,51 Some studies highlighted concerns regarding the long-term sustainability and scalability of AI systems. Their value often depended on continuous data entry with clinical workflows, which had been inconsistently addressed in many implementations.^33,39 Other studies noted that the absence of external validation across populations further undermined confidence in the broader applicability of AI in models trained datasets.^40,53 Importantly, the perception of value was shaped not only by clinical performance but also by social, ethical, and psychological factors. The studies highlighted that burnout among providers, concerns about professional displacement, or doubts about fairness reduced enthusiasm for AI, even when technical performance was satisfactory.^42,43,49 Conversely, when users perceived the AI as helpful or enjoyable, their engagement and continued use increased.^41,47

Domain 4: The adopter system

The adopter system domain focuses on the attitudes, capacities, and behavioral responses of individuals expected to engage with the technology, including staffs, patients, and caregivers. ³¹ Successful AI adoption in clinical settings was often influenced by clinicians’ perceived usefulness and trust in the system. Studies reported that physicians in primary settings, expressed openness to AI-assisted decision-making, especially when it enhanced diagnostic accuracy or reduced workload.^34,42 Adoption was also supported by practical experience, and peer influence between AI tools and clinical needs. For instance, continuous exposure, endorsement by colleagues, and the presence of psychological safety increased clinicians’ willingness to use AI tools.^35,37,47

Trust has become a factor in promoting and discouraging adoption among user groups. Models with higher interpretability or those integrated into familiar platforms fostered trust among both clinicians and patients.^34,41,45 However, adoption was frequently hindered by fears of skill degradation and professional displacement among experienced practitioners.^46,48 Emotional and psychological factors such as anxiety and a perceived lack of control further discouraged adoption, even when the perceived utility was high.^42,43

Among patients and caregivers, adoption varied by age, education, and digital literacy. Younger users with higher levels of technological proficiency, and those with prior positive experiences with AI, demonstrated higher levels of acceptance, whereas older adults or individuals with limited technological familiarity typically required additional training or support.^36,41,55

Several studies emphasized the importance of structured workforce training for AI adoption. Training needs were identified across multiple roles, including physicians, nurses, and diagnostic technicians.^35,36,41 In particular, frontline clinicians in primary care and junior medical staff often expressed lower confidence in AI tools due to limited exposure or prior experience.^37,39 Content areas for training included not only technical use and troubleshooting but also data ethics, bias awareness, and interpretation of AI outputs. For instance, Zeng et al. ³⁶ proposed a training model, offering differentiated instruction to medical staff, supplemented by digital feedback via messaging apps like WeChat.

Domain 5: The organization

The organization domain assesses an organization's ability to adopt innovation, including its change readiness, funding mechanisms, required workflow adjustments, and the implementation efforts associated with new technologies. ³¹ Successful implementation of AI in clinical settings often depended on institutional support and technical infrastructure. Several studies indicated that hospitals with centralized leadership, dedicated AI teams were more likely to facilitate adoption.^34–36,46 Scalability was further supported when AI systems aligned with existing hospital workflows or could be built upon current platforms, such as EMRs or remote monitoring systems. For example, some studies indicated that using existing government or digital health frameworks helped reduce deployment costs and enhance feasibility in both urban and community settings.^43,55

In addition to technical aspects, organizational management capacity and interdepartmental coordination also influenced AI adoption. Institutions that fostered a learning-oriented culture that support innovation were more open to piloting AI tools and refining their implementation.^34,44,47 Hospitals with formalized mechanisms for change management, such as structured training programs, demonstrated stronger adaptability in AI adoption. Pilot testing was commonly used as a strategy to mitigate staff resistance prior to implementation. For example, AI tools were initially deployed in piloted departments to allow time for stakeholder feedback and process adjustments. ⁵⁴ Moreover, collaboration between clinical, technical, and administrative units proved beneficial, particularly in studies involving hospital systems, or AI-supported screening across multiple provinces.^45,53,54

Despite facilitators, several organizational barriers were reported. First, technical readiness varied across institutions. While large hospitals demonstrated the infrastructure and human resources to support AI deployment, smaller hospitals and rural clinics often lacked IT systems, network capacity, or technical expertise.^33,46,58 Second, unclear internal governance structures weakened institutional capacity for scale-up. In some settings, hospitals lacked formal workflows for AI evaluation, making to fragmented implementation and reduced accountability.^44,48,54 Third, resource-limited facilities expressed concern about the sustainability of AI systems, particularly when deployment lacked training or support. In such settings, the absence of continuous professional development led to inconsistent follow-up into daily workflow.^45,55

Domain 6: The wider context

The wider context domain encompasses the political, regulatory, professional, and sociocultural forces that shape the feasibility and acceptability of AI implementation in clinical practice. ³¹ In many studies, national policy initiatives played a role in facilitating AI adoption. Government funding programs provided both financial and strategic support.^33,34,46,50 This was particularly evident during the COVID-19 pandemic, when the urgency of the crisis accelerated the deployment of digital tools. ³³ Broader policy trends, such as the promotion of smart hospitals and digital transformation, also contributed to an environment for both research and clinical experimentation.^41,42,44,47 Governmental alignment also proved essential in efforts, as seen in coordinated public health deployments involving hospitals, health authorities, and community organizations.^43,54

However, barriers still persisted around data governance, legal responsibility, and ethical oversight. Regulatory uncertainty was a recurring concern, with many studies highlighting the absence of clear guidelines on data privacy and the legal status of AI recommendations in clinical settings.^{38,43,48,54,55} Ethical risks and unclear liability frameworks further constrained technology acceptance. Studies reported that users feared unauthorized data access without clear responsibility attribution.^43,44,48 Moreover, social and economic difference influenced AI acceptance. Regional differences in digital infrastructure and digital literacy at population level created uneven conditions for implementation in resource-limited areas.^45,58

Domain 7: Embedding and adaptation over time

The domain of embedding and adaptation over time examines how technologies evolve, scale, and sustain within clinical environments beyond implementation. ³¹ In the context of AI in healthcare, successful long-term integration was often associated with system refinement and feedback loops. Several studies highlighted the importance of real-world data accumulation in enhancing model performance over time. Systems designed with updated architectures demonstrated greater adaptability and long-term relevance in care environments. For example, some systems enabled retraining with routine clinical data or supported systemic upgrades.^34,36,45,55 Collaborative arrangements between healthcare institutions and technology providers also played a role in supporting long-term integration. Building on the technical adaptability described above, organizational partnerships helped ensure that AI systems evolved in ways that aligned with clinical workflows. For instance, when IT personnel were embedded in care teams with clinicians, systems demonstrated a higher likelihood of refinement and relevance.^35,46

Nonetheless, several barriers were there to sustained integration of AI technologies. Many tools remained at the pilot stage, with limited evidence of full incorporation into routine clinical workflows.^33,37,39 The absence of mechanisms weakened the ability to assess real-world effectiveness over time.^54,55 In some cases, tools designed for short-term crisis response lacked plans for institutionalization, limiting their long-term utility.^43,57 In addition, the lack of clear governance structures for long-term AI maintenance also posed challenges to sustained deployment. Without standardized practices for embedding AI into clinical education, IT infrastructure, and quality improvement, systems failed to transition from one-time use to ongoing service.^48,49,56

Key facilitators and barriers to utilizing AI in clinical practice

Key facilitators and barriers were synthesized from thematic modes observed across the NASSS domains, with each domain contributing unique aspects influencing AI utilization. Figure 4 presented six key facilitators identified through qualitative synthesis of the included studies, guided by the NASSS framework. These facilitators reflected enablers that were linked to implementation outcomes across 1 or more of the seven NASSS domains. Specifically, six key facilitators were identified, that were, strong perceptions of increased operational efficiency, availability of AI tools that supported clinical decision-making, confidence in improved diagnostic accuracy, alignment with personalized care goals, perceived cost-saving potential, and enabling environments shaped by policy, governance, and institutional commitment.

Figure 4.

Key facilitators of utilizing AI in clinical practice and mapped to relevant domains of the NASSS framework.

In contrast, seven key barriers were commonly reported, that were, ethical and privacy concerns, limited user acceptance among clinicians or patients, inconsistent or unvalidated accuracy in real-world settings, technical complexity requiring substantial training, unclear accountability for AI-driven decisions, trust-related issues, and inadequate infrastructure in low-resource environments. Figure 5 summarized the seven key barriers as synthesized from the included studies and mapped across the NASSS framework domains.

Figure 5.

Key barriers of utilizing AI in clinical practice and mapped to relevant domains of the NASSS framework.

Discussion

Principal findings

This review contributes valuable insights into the utilization of AI in clinical practice, with a focus on China's experiences. It offers a broader understanding of AI implementation beyond applications limited to specific diseases or clinical specialties. A finding is the fragmented nature of existing research, which often centers on narrow clinical or operational aspects while overlooking the systemic complexities that influence successful AI adoption. This sits in contrast to the increasing global enthusiasm for AI integration in healthcare systems, particularly in ethnic Chinese communities worldwide and other developing regions. The review further reveals a lack of comprehensive frameworks for evaluating the multifaceted outcomes of AI implementation across diverse AI types. By synthesizing outcomes across included studies and identifying both key facilitators and barriers, this review moves beyond conventional evaluations and provides a more comprehensive foundation for future research and implementation efforts, especially in settings that share similar demographic and resource challenges with China.

This review highlighted the potential of various AI technologies to enhancing diagnostic accuracy, workflow efficiency, patient engagement, and cost-effectiveness. Whether as a complement to existing clinical practices or as a transformative force on its own, AI technologies showed tremendous potential in advancing clinical practice.^{33,34,36,38,42,49,52,58} Findings from this review further suggested that AI-assisted approaches performed at least as well as conventional methods, and in some cases even better. Several studies provided comparisons between AI-assisted approaches and conventional clinical methods. Where data were available, reported metrics included accuracy, sensitivity, specificity, and time efficiency. For example, Zhang et al. ³³ reported that a AI system achieved a diagnostic accuracy of 92.49%, sensitivity of 94.93%, and specificity of 91.13% in distinguishing novel coronavirus pneumonia, surpassing conventional radiological interpretation. Similarly, Q. Liu et al. ⁵³ argued that AI-based bone age assessment achieved high accuracy on both internal (97.7%) and external (89.5%) test sets, which in some cases exceeded the consistency of traditional radiologist assessments. Moreover, Li et al. ³⁴ demonstrated that the implementation of an AI triage system led to a significant reduction in patient waiting times, with improvements in service efficiency compared to conventional manual workflows.

Different AI technologies demonstrated distinct roles across various outcome dimensions. For example, ML mainly played a role in enhancing diagnostic accuracy and clinical efficiency, NLP showed advantages in patient communication, knowledge delivery, and follow-up management, while DL and CV exhibited potential in medical image analysis, workflow optimization, and cost containment. A deeper examination of methodological quality revealed differences across different types of AI applications. Studies involving DL and CV generally received higher methodological ratings. This might be attributed to their consistent use of structured clinical datasets and validated performance metrics such as area under the receiver operating characteristic curve, sensitivity, and specificity, as well as the frequent adoption of retrospective cohort designs. ⁴⁰ In contrast, studies utilizing NLP, particularly those focused on generative AI tools such as ChatGPT, tended to be exploratory in nature, were often based on small sample sizes, and lacked external validation.^56,57,59,60 ML-based studies also demonstrated mixed methodological rigor, with variations in reporting standards and evaluation frameworks. Bias assessment conducted in this review further reflected these differences. In the assessment of bias, studies of DL and CV were more frequently rated as low risk of bias, while NLP and ML studies were more often rated as having moderate^{37,39,47,50,53,58} or high risk, ⁵¹ these may be due to limitations in study design, outcome assessment, and sample representativeness. These variations may affect the comparability and generalizability of findings across AI domains, and should be considered when interpreting the strength of evidence.

Findings from this review identified six key facilitators of utilizing AI in clinical practice, regarding strong perceptions of increased operational efficiency, availability of AI tools that supported clinical decision-making, confidence in improved diagnostic accuracy, alignment with personalized care goals, perceived cost-saving potential, and enabling environments shaped by policy, governance, and institutional commitment. Findings also identified seven key barriers, such as ethical and privacy concerns, limited user acceptance among clinicians or patients, inconsistent or unvalidated accuracy in real-world settings, technical complexity requiring substantial training, unclear accountability for AI-driven decisions, trust-related issues, and inadequate infrastructure in low-resource environments. Notably, diagnostic accuracy emerged as both a facilitator and a barrier across the included studies. On the one hand, advanced algorithms were shown to significantly enhance diagnostic precision compared to conventional clinical approaches. ³³ On the other hand, there were concerns about algorithmic bias and lack of transparency in the AI decision-making process due to data quality. ⁵⁶

The results of this review were partly consistent with previous studies. Sharma et al. ⁶¹ argued that AI systems played a critical role in clinical care, particularly by optimizing patient-provider encounters and improving the quality of care. Similarly, Kee et al. ⁶² emphasized the contribution of ML in clinical prediction, they illustrated its potential in developing prediction models to assess the risk of cardiovascular disease among patients with type 2 diabetes. While these studies highlighted AI's potential, Wubineh et al. ¹⁹ pointed out several challenges that hindered the widespread adoption of AI in healthcare, including ethical and privacy concerns, lack of awareness, technological unreliability, and professional liability. In contrast, Lee and Yoon ⁶³ presented a more optimistic view, outlining the numerous opportunities AI offers, such as improved disease treatments, enhanced patient engagement, reduced medical errors, better service quality, increased operational efficiency, cost reduction, productivity gains, and new job creation. This review highlighted the tremendous potential of AI to enhance diagnostic accuracy, workflow efficiency, patient engagement, and cost-effectiveness, while also assessing the key facilitators and barriers to AI utilization within a comprehensive framework. Despite growing enthusiasm for AI, this review identified a lack of evidence evaluating the cost-effectiveness of AI implementation in real-world clinical settings. While some studies reported that AI may reduce costs through increased efficiency ³⁹ and automation, ⁴⁷ few included formal economic evaluations or long-term assessments of financial impact. Moreover, most reported improvements were limited to intermediate outcomes, such as task efficiency or workflow optimization, ⁴⁶ without linking these gains to health system savings or patient-centered outcomes, such as morbidity, mortality, or quality of life. This pointed the need for future research to incorporate robust economic and clinical evaluations in order to guide sustainable and value-based integration of AI into healthcare systems.

Key characteristics

This review was one of the first systematic evaluations to examine the utilization of AI in clinical practice within the China's context, and differed from the existing literature that mainly explored AI applications in Western contexts. Our analysis highlighted several characters that influenced AI adoption in China. Firstly, the sociocultural acceptance and integration of AI in China have been facilitated by government initiatives. Among these, the Healthy China 2030 plan ⁶⁴ established national priorities and allocated resources specifically aimed at promoting AI adoption across the healthcare system. ⁶⁵ Government advocacy and strong policy support had cultivated a favorable climate, enhancing acceptance among healthcare providers and institutions, and thus facilitating AI's incorporation into existing clinical workflows. The high level of trust in institutional authority within Chinese society made governmental endorsement a crucial factor in shaping both clinicians’ and patients’ willingness to adopt AI technologies.

Secondly, China's substantial demographic scale provided a distinct advantage by generating vast and diverse clinical datasets sourced from a varied population base. These datasets accelerated algorithm refinement and improved AI performance adaptability across different demographic groups, thereby enhancing clinicians’ confidence in adopting these technologies. However, demographic diversity also introduced notable challenges. While AI contributed to expanding access in underserved areas, significant variations in digital literacy, particularly in rural regions and among older populations, often limited the effective use of AI tools. It highlighted the need for training and system design. In addition, differences between urban and rural regions further influenced AI adoption. AI-driven telemedicine and smart hospitals played a key role in addressing these inequities by improving access to clinical services in rural and remote areas.^53,66,67 This ability to bridge healthcare divide demonstrated a specific adaptation of AI technologies that was responsive to the demographic realities of China or other countries with similar healthcare challenges.

Thirdly, sociocultural factors also included the integration of AI with traditional Chinese medicine (TCM). AI-driven TCM applications effectively capitalized on the cultural familiarity and acceptance of TCM within the Chinese population. Consequently, these applications resonated with healthcare providers and patients, greatly facilitating their adoption. Such integration promoted cultural acceptance among global ethnic Chinese communities who share similar medical traditions and cultural values. ⁶⁸ Nevertheless, cultural preferences for personalized care in China, such as the reliance on personal interaction, might lead to skepticism among certain population groups. This skepticism toward replacing traditional physician roles with automated systems might hinder the adoption of AI in specific clinical contexts within China.

The integration of AI into clinical practice in China was influenced by specific sociocultural and demographic contexts, including governmental policy support, demographic advantages, and cultural alignment. Additionally, factors such as public trust in institutional authority, disparities in digital health literacy, demographic diversity between urban and rural regions, the needs of an aging population, and cultural preferences for human-led care also shaped the patterns of AI adoption in China. However, many of these enabling conditions were embedded within China's unique political and regulatory system. China's governance structure, characterized by state coordination and policy driven planning, had contributed to favorable conditions for the large scale and timely adoption of AI technologies.^34–36,46 While effective in driving implementation, such a top-down approach might reduce responsiveness to local needs and constrain contextual customization, particularly in decentralized healthcare systems. It might reduce local adaptability, inhibit bottom-up innovation, and lead to reduced flexibility when clinical needs and technologies evolve rapidly. In addition to structural influences, ethical or legal issues remained underdeveloped in the Chinese healthcare context. Concerns included ambiguous standards for data privacy protection, the absence of explicit informed consent protocols for AI-supported care, and fragmented data governance frameworks.^{38,43,48,54,55} In contrast to structured frameworks such as the European Union's General Data Protection Regulation, ⁶⁹ China's regulatory approach to health data remains relatively nascent. These might limit transparency of AI-driven care, thereby potentially affecting trust and the sustainable integration of AI technology in clinical practice. Still, China's experience offered valuable insights for other developing countries and communities with similar cultural backgrounds and demographic challenges. It provided a reference point for shaping the adoption of AI in clinical practice.

Recommendations and future directions

Findings from this review highlighted that, although AI offered significant advantages, these needed to be carefully weighed against a range of barriers as identified in this study. To address identified barriers from this review, several practical solutions had been proposed. For example, promoting the use of explainable AI models could enhance transparency and trust, while algorithm validation using diverse, representative datasets could reduce bias and improve performance consistency.^70,71 Tailored training programs for clinicians were essential to overcome technical barriers and improve digital competence. ⁷² Furthermore, robust governance frameworks, including clear lines of accountability and ethical oversight, were necessary to support responsible AI deployment. ⁷³ Comparative findings from the reviewed studies suggest that while AI could outperform conventional methods in specific contexts, its successful integration depended on local infrastructure readiness, regulatory alignment, and clinician engagement. ⁷³ These highlighted the importance of developing multidimensional strategies that address the factors limiting AI adoption. To this end, the review proposed five key recommendations that respond to the identified barriers and integrate relevant contextual considerations. These recommendations aimed to bridge the gap between technological development and real-world clinical implementation. While informed by evidence from China's experience, the applicability of these recommendations in other contexts might require contextual adaptation. Many of the enabling conditions, such as national policy support, government investment, and institutional coordination, were rooted in China's centralized healthcare governance system. In contrast, health systems with more decentralized governance structures, limited regulatory capacity, or differing political and cultural environments might face challenges in applying comparable strategies. Stakeholders are therefore encouraged to critically evaluate the relevance of these recommendations and adapt them to align with local infrastructure, regulatory frameworks, and societal expectations. The recommendations were summarized in Textbox 1.

Textbox 1.

Summary of key recommendations for future directions

1. Close evidence gaps

   • Conduct long-term clinical evaluations of AI systems to generate rigorous evidence on patient outcomes and cost-effectiveness.
   • Investigate patient and clinician acceptance across different healthcare settings to inform deployment.
   • Apply implementation science to understand how contextual factors, such as workflow integration, institutional support, or digital maturity, affect adoption.
2. Improve usability, transparency, and trust

   • Develop explainable AI tools with clinicians to improve real-world usability.
   • Embed AI tools into clinical workflows with real-time decision support and culturally informed interface design.
   • Provide scenario-based training programs for clinicians to build confidence in AI-assisted care.
3. Address ethical risks and data governance

   • Implement data-sharing protocols, such as secure data enclaves, to ensure both access and protection.
   • Encourage stakeholder (patients, frontline clinicians, and marginalized populations) engagement during AI development.
4. Strengthen policy and infrastructure readiness

   • Provide funding and infrastructure support for pilot implementation in underserved settings.
   • Establish legal frameworks that clarify accountability and responsibility for AI-assisted decisions.
5. Standardize evaluation and oversight mechanisms

   • Develop clear evaluation protocols for emerging technologies, such as NLP and multimodal AI, before and after deployment.
   • Support model monitoring through shared governance structures for updates, retraining, and continuous quality assurance.

Limitations

This study has several limitations. Firstly, publication bias is long-standing and recognized phenomenon in health research ⁷⁴ and may have influenced the broadly positive results in this study. Beyond this, structural factors within the Chinese research and publication environment may have further contributed to this imbalance. Studies that align with national priorities or report successful implementation outcomes are more likely to receive funding, approval, and publication support, while studies that reveal challenges or failures may be underreported or unpublished. This bias may have contributed to a tendency toward overrepresentation of favorable outcomes in the evidence base included in this study. Secondly, this study applied the JBI and NOS tools in assessing included studies. Although both tools are widely accepted, their application may involve subjective judgment, particularly in complex cases where interpretation of criteria may vary across researchers. Thirdly, potential selection bias in the included studies should be noted. The heterogeneity in study populations, AI applications, clinical settings, and evaluation methods might have limited the generalizability of the findings. Notably, many studies were conducted in urban hospital environments, which might not have represented broader healthcare contexts. Furthermore, variation in AI development levels, data sources, and outcome measures contributed to methodological heterogeneity, potentially affecting the consistency and comparability of the results. Fourthly, to ensure consistency in search strategy and focus on peer-reviewed sources with broader global impact, this review restricted its search to three English-language databases. As a result, relevant articles from other databases, particularly Chinese-language sources such as China National Knowledge Infrastructure and the Chinese Academy of Sciences Library, might have been excluded. Finally, while this review initially considered studies published between 2010 and 2024, the earliest eligible publications included were from 2020. This could be attributed to the recent application of AI in China's healthcare sector, which was likely driven by technological advances by the COVID-19 pandemic. A strength of this review was the use of a systematic search strategy, which enabled the identification of relevant studies despite limitations in quantity and scope. Moreover, the analysis of this review was informed by comprehensive frameworks, adding conceptual depth and coherence to the synthesis.

Conclusions

This review emphasized the tremendous potential of AI utilization in clinical practice. AI implementation has been associated with improvements in diagnostic accuracy, workflow efficiency, patient engagement, and cost-effectiveness. Evidence from the included studies suggests that AI-assisted approaches perform at least as well as conventional methods, and in some cases, even better. This review also identified key facilitators that support the use of AI in clinical settings, although several barriers remain. In addition, key characteristics of AI adoption within specific contextual settings were synthesized, and contextually informed recommendations were proposed to facilitate its integration and address the identified barriers. However, the long-term impact of AI integration on clinical outcomes was beyond the scope of this study. Future research should focus on evaluating the sustained effects of AI in routine clinical practice, developing implementation frameworks to support its long-term adoption, and addressing emerging issues as AI becomes increasingly embedded in healthcare systems.