This review was reported following the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 statement [22]. The protocol was registered with PROSPERO in October 2024 (registration no CRD42024598884). A theoretically grounded mixed methods systematic review provides a structured and integrative approach to identifying the barriers and facilitators that influence patient acceptance of AI.

The literature search was conducted following the PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Searching) reporting guidelines [23]. Two researchers (HS and JY) and an evidence-based expert (JH) developed the search strategy using the Population, Interest, and Context framework [24]. Eligibility criteria focused on patients, AI artifacts, and attitudes and perceptions. Barriers and facilitators were defined as factors preventing or motivating AI use. The search strategy for this study was built around the core terms “artificial intelligence,” “patient,” and “attitude,” combining the use of medical subject headings and keywords. A comprehensive search was conducted across PubMed, Web of Science, Embase, Cochrane Library, CINAHL, CNKI, Wan Fang, VIP, and SinoMed (inception to December 23, 2025). Detailed search strategies are provided in Table S1 in Multimedia Appendix 1. Our comprehensive search strategy was executed across multiple sources to ensure literature saturation. This included searching major study registries, systematic online database searches, manual browsing of key journals and organizational websites, and citation searching (both backward and forward) of all included studies and relevant reviews. No restrictions were applied based on publication status or language during the initial search phase.

We included qualitative, quantitative, and mixed methods studies that addressed (1) patients’ perceptions of using AI in health care (diagnosis and treatment) and (2) barriers and facilitators to patients’ willingness to use AI. Studies limited to health care professionals, managers, technology developers, and medical and nursing students, as well as studies that only tested the effectiveness of AI, were excluded. Table 1 shows the inclusion and exclusion criteria for this study.

Two researchers (HS and MH) independently screened papers after removing duplicates in EndNote (version 21; Clarivate), resolving disagreements through discussion or third-party consultation (JH). Data on barriers, facilitators, and interventions were extracted using a data extraction table [24]. Qualitative studies provided themes and participant quotes, while quantitative studies included outcome measures and narrative summaries. Mixed methods studies were analyzed separately. Extracted details included author, year, design, participant characteristics, findings, and interventions, with open-ended survey responses treated as qualitative data, ensuring comprehensive extraction and synthesis. When necessary, corresponding authors were contacted to obtain missing or supplementary data.

The Mixed Methods Appraisal Tool (MMAT) [25] was used to assess the methodological quality of the included studies, covering qualitative, quantitative (randomized/nonrandomized), and mixed methods designs. Two researchers (HS and MH) independently conducted assessments, resolving disagreements through consensus or consultation with a third expert (JH).

A comprehensive search of 9 electronic databases using a refined search strategy identified 7452 records. After removing duplicates, 6680 records remained. Following an initial screening of titles and abstracts, 128 records were deemed relevant. Subsequently, full-text screening was performed according to the inclusion and exclusion criteria, excluding irrelevant papers. Ultimately, 61 studies were included in the final review. All literature was used to extract barriers and facilitators, with 53 studies used to extract potential interventions. The literature screening process is shown in Figure 1.

This study included 61 studies categorized by research design: 20 qualitative (primarily semistructured interviews), 2 randomized controlled trials, 6 mixed methods, 6 quantitative descriptive, and 27 nonrandomized studies. Quantitative studies mainly used questionnaires, with one adapting the UTAUT2 model [36]. Sample sizes ranged from 9 to 2899 participants. Geographically, studies spanned 19 countries, with 45 from 15 developed countries (eg, 15 from the United States) and 16 from 4 developing countries (eg, 9 from China). Detailed characteristics of the included studies and participants are summarized in Table 2.

The study population includes patients using medical AI aged ≥18 years, recruited from communities, hospitals, and online surveys. Patient types varied, with a focus on patients with cancer (n=11), radiology (n=8), perioperative (n=4), diabetes (n=3), dental (n=5), and stroke (n=2). Nine studies included patient and health care professional perspectives; only patient data were analyzed to align with research objectives. The comprehensive data and full extraction details are provided in Table S6 in Multimedia Appendix 1 due to space constraints.

According to the MMAT evaluation, the average score of the included studies was 4.4. Approximately 85% (n=53) of the studies met all or 80% of the quality assessment criteria. About 11% (n=6) of the studies met 60% of the quality assessment criteria. Two studies met only 40% or less of the quality assessment criteria. In general, the methodological quality of the included studies was assessed as relatively high. The MMAT User Manual mentions the treatment of low-quality studies, which are rounded off according to the topic of the study; in this study, these 2 papers were valuable in contributing to the research topic and were therefore not excluded [25]. The MMAT scores for each included study are provided in Table 2, and detailed item-level assessments are provided in Table S7 in Multimedia Appendix 1.

Through group discussions, 27 factors from UTAUT2 and TDF domains were mapped to 93 potential BCTs. After consolidation, 25 distinct BCTs were assessed using the Affordability, Practicability, Effectiveness and Cost-Effectiveness, Acceptability, Side-effects or Safety, and Equity criteria and categorized by feasibility: 6 (24%) as high, 14 (56%) as medium, and 5 (20%) as low (detailed in Table S10 in Multimedia Appendix 1). High-feasibility BCTs (eg, 4.1 and 5.1) are low cost, easy to integrate, and well accepted. Medium-feasibility BCTs (eg, 6.1 and 15.1) are evidence-based but face operational constraints, while low-feasibility BCTs (eg, 9.1 and 12.1) require system- or policy-level changes beyond clinical teams. Concrete examples for these BCTs were drawn from the literature (Table S9 in Multimedia Appendix 1), resulting in 40 targeted interventions. Complete results are shown in Table 4 (the serial numbers and other details of the relevant BCT are provided in Table S3 in Multimedia Appendix 1).

This review aimed to explore the main barriers and enablers underlying patients’ adoption of AI in hospital settings. Prior studies have examined the intention to use AI among different types of patients, but there has been no systematic synthesis of barriers and facilitators. This study used an innovative approach to synthesize patients’ views on using AI in medicine, integrating barriers and facilitators of patient adoption of AI in medicine based on the UTAUT2 and TDF.

Our findings reveal that AI in medicine adoption is driven by a 3-tier mechanism: “technology acceptance, contextual modulation, and ethical security.” Moreover, individual differences influence AI acceptance. For example, age moderates the perception of technological complexity, with older adults favoring simpler systems [67,72]. Education level is a cognitive mediator, affecting performance expectancy by shaping AI-related knowledge [51,60,65,71,75,84]. Additionally, disease severity exhibits an inverted U-shaped relationship with AI acceptance, where patients with moderate conditions demonstrate the highest acceptance [64,73].

Regarding the most important barriers and enablers identified in this review, performance expectancy and effort expectancy were the most cited factors influencing patients’ use of AI in medicine. This finding substantiates the results of earlier studies [37,61,67]. Patients’ recognition of AI’s diagnostic efficiency and accuracy emerges as the primary determinant of acceptance under performance expectancy [15,30], aligning with findings from Schmitz et al [16] in cross-national telemedicine research. Additionally, within the TDF framework, the “knowledge” domain significantly impacts patient adoption of AI, highlighting that awareness and understanding of AI technologies are crucial determinants of acceptance and use [37,61,67]. There is a need to improve AI knowledge and increase willingness to use AI, which can be achieved by enhancing patient education through outpatient health seminars and multimedia educational videos.

In addition, the operational feasibility of AI plays a key role in promoting patient use. The perceived value of AI for patients increases significantly when they are able to receive positive feedback and are effectively supported by health care professionals when using it [95]. For example, an AI robotic system that receives positive feedback from health care professionals during rehabilitation training can enhance patients’ confidence and sense of accomplishment, further promoting the usage of AI [96]. However, poorer usability can be a source of fear for patients toward new technologies and, therefore, may hinder their adoption [96]. A “progressive complexity” strategy, embedding educational modules within basic applications, may facilitate gradual comprehension and trust development in AI systems [36,84].

Social networks profoundly shape patient willingness to use AI, extending beyond physician recommendations to include family members, friends, and community influence. Family acceptance acts as a critical multiplier—tech-savvy relatives often facilitate AI adoption by assisting with device operation or interpreting digital health reports, whereas households with low health literacy may resist AI due to distrust of unfamiliar technology [63]. Peer influence also plays a role: patients are more likely to embrace AI-driven tools after observing positive experiences shared by friends or online communities, particularly in managing chronic conditions such as diabetes or hypertension. Conversely, negative anecdotes within social circles (eg, stories of AI misdiagnosis) can amplify skepticism, creating “information cascades” that undermine trust [47]. These dynamics highlight the need for targeted social marketing strategies, for example, family-centered education workshops or peer advocacy programs to address intergenerational and community-specific barriers.

Cultural adaptability is primarily shaped by specific constructs within the UTAUT2 and TDF frameworks, with social influence being the most culturally sensitive dimension. This review finds that in collectivist societies (eg, East Asia), social influence often manifests as deference to family or group consensus [97], effectively shifting the adoption unit from the individual to the collective. Conversely, in individualistic cultures, it leans more on trust in professional authority or empirical evidence. Consequently, implementation strategies must be contextualized: collectivist settings may require designs that facilitate family-inclusive decision-making and leverage community opinion leaders, whereas individualistic contexts should emphasize transparent personal benefits and institutional endorsements. Systematically analyzing culture through these theoretical lenses enables more targeted and effective AI implementation across diverse health care systems.

Ethical safety concerns, particularly the “algorithmic black box,” undermine trust in AI’s reliability among patients and clinicians [45,58,62,64,67]. Explainable AI, by providing transparent diagnostic reasoning and interpretable reports, mitigates skepticism and strengthens technical credibility [39]. While AI may reduce patient embarrassment in sensitive consultations [46], overreliance risks eroding humanistic doctor-patient interactions, exacerbating anxiety [37,40,44,45,49,50]. Balancing privacy protection with human-centered design remains a critical research frontier.

Beyond mitigating ethical and relational barriers, our synthesis points toward the necessity of proactively rehumanizing AI-driven health care. This concept involves intentionally designing AI systems and their implementation contexts to preserve and enhance, rather than replace, essential human elements such as provider presence, empathy, and meaningful interaction. Studies, such as the work of Almokdad et al [98] in digital hospitality, underscore the critical importance of human presence in mitigating the impersonal nature of automated services. Translating this to health care, AI should be positioned as a tool that augments, rather than replaces, the clinician. To operationalize this, concrete strategies include (1) adopting “AI-as-assistant” design paradigms where clinicians remain central in communication and decision-making, (2) implementing empathy-aware algorithms to prompt providers to address patient emotional states, and (3) leveraging AI’s efficiency gains to protect or increase time for direct patient-clinician interaction.

Additional psychological and economic factors include hedonic motivation, habituation, and price-value trade-offs. Positive emotional experiences with AI (eg, intuitive interfaces) boost initial acceptance [36,47,83,89,96], whereas habit-driven preference for traditional methods necessitates proactive demonstration of AI’s clinical advantages [38,62,66,83,87,89]. Although AI adoption may incur higher upfront costs [9,58,63,66,67,81,89], its long-term potential to reduce unnecessary diagnostics and optimize treatment pathways could yield cost savings [88,90]. Future research should contextualize AI’s economic impact within health care settings while prioritizing personalized, culturally sensitive implementations.

The BCT taxonomy (version 1) provides a standardized framework for behavioral interventions in AI implementation [34]. This study advances its practical application by introducing a feasibility lens, which transforms a theoretical list of techniques into a prioritized implementation strategy. Our assessment distinguishes high-feasibility BCTs (eg, instruction and goal setting) as immediately actionable tools for frontline adaptation, such as optimizing interfaces for older adults [99]. Medium-feasibility BCTs, including strategies such as social comparison to address privacy concerns [55], require deliberate planning and resources. Ultimately, sustainable adoption depends on low-feasibility, systemic enablers, such as transparent data governance, which establish the necessary foundation for trust. Integrating barrier mapping with this feasibility-driven prioritization creates a clear, staged model for implementation. It clarifies what clinicians can do now, what requires organizational support, and what must be advocated at the policy level, offering a pragmatic roadmap for integrating AI into health care.

In summary, the primary contribution of this study lies in its theoretical integration and methodological innovation. Distinct from previous research that predominantly focused on technological attributes or isolated psychosocial factors, this review establishes a comprehensive analytical model that links technology acceptance with behavior change by integrating the UTAUT2 and TDF frameworks. Beyond this theoretical contribution, this integrative model deepens the understanding of barriers and facilitators to patient adoption of AI and, importantly, translates these insights into actionable evidence by mapping the findings to specific and prioritized BCTs. This evidence directly informs and empowers multiple stakeholders: it provides health care administrators with a blueprint for implementation strategies, equips clinicians with tools for patient engagement, guides AI developers toward human-centered design principles, and supports policymakers in developing frameworks for ethical evaluation and governance. Ultimately, this review bridges the gap between theoretical understanding and practical application, thereby enhancing the translational value of research on AI acceptance in health care.

The strength of this study lies in its systematic synthesis of qualitative and quantitative evidence, conducted through a mixed methods evaluation guided by an integrated theoretical framework. Theoretically, it innovatively combines the UTAUT2 model with the TDF to analyze barriers and facilitators influencing patients’ adoption of AI in health care, thereby providing a dual perspective that incorporates both technology acceptance and behavioral science. This theory-informed approach allows for a coherent interpretation of existing evidence and establishes a robust foundation to support patient-centered AI implementation, while also informing the design of future interventions and health policy initiatives.

This review has several limitations. First, methodological constraints include potential publication bias and language restrictions (Chinese and English only), which may affect the comprehensiveness and cultural generalizability of the findings. Second, the evidence base is limited by the uneven geographical distribution of included studies, with underrepresentation from developing countries, and the retention of 2 lower-quality studies that did not alter the overall conclusions. Third, the proposed interventions are theoretically derived from TDF-BCT mapping and lack empirical validation. Finally, given the rapid evolution of AI technology and its regulatory context, the findings represent a snapshot in time and require ongoing re-evaluation as the field advances.

This study uses a mixed methods approach to explore key factors influencing patient acceptance of health care AI. Integrating the UTAUT2 model with the theoretical domains of the TDF systematically identifies barriers and facilitators, emphasizing the need for tailored implementation strategies. Theory-based interventions are proposed by mapping the TDF to BCTs. Findings highlight that building trust in technology and prioritizing human-centered design are essential for acceptance, while privacy and ethical concerns remain significant barriers. Further research is needed to assess the effectiveness and sustainability of patient-centered, multicomponent interventions and to establish a dynamic monitoring system for evolving patient perceptions amid technological advancements.