Virtual patients are sophisticated educational simulations designed to replicate authentic clinical scenarios, enabling health care trainees to practice skills in a safe, controlled environment without risk to real patients [1,2]. They are defined as “a representation of an actual patient,” which can include various forms such as software-based simulators or manikins, and specifically as “a computer program that simulates real-life clinical scenarios in which the learner acts as a health care provider,” making clinical decisions [3]. They serve as a cornerstone of modern clinical education, aiming to enhance clinical reasoning, communication proficiency, and decision-making abilities [4]. The recent and rapid integration of generative artificial intelligence (GenAI)—a subset of artificial intelligence (AI) powered by large language models (LLMs) and natural language processing—is fundamentally transforming this educational tool [5,6]. Unlike traditional scripted simulations, GenAI-supported virtual patients can generate dynamic, adaptive, and contextually relevant responses in real-time, allowing for more realistic conversations, emotional responsiveness, and personalized learning experiences [7,8]. This shift marks a significant evolution in simulation-based learning, moving from static, preprogrammed cases toward interactive, intelligent patient encounters.

The evolution of virtual patients demonstrates a clear trajectory toward greater interactivity and realism. From early, static computer-based cases on systems such as PLATO [9], modern iterations now incorporate multimedia, branching narratives, and data-driven models to create more engaging and realistic clinical encounters [10,11]. This evolution has yielded significant educational benefits. Empirical studies show that virtual patients can improve clinical decision-making, diagnostic accuracy, and foundational skills such as screening and referral across various health disciplines [11,12]. For instance, they have been successfully implemented as standardized, unfolding simulations to replace scarce pediatric clinical hours while maintaining clinical competency in nursing education [13]. Furthermore, a pilot study grounded in Experiential Learning Theory demonstrated that virtual patient simulation led to statistically significant improvements in clinical reasoning and communication skills among prelicensure nursing students [14]. They provide a scalable, consistent training environment that addresses limitations inherent to human standardized patients, such as fatigue and variability [7,15].

Despite these advances, a critical constraint remains: the predominant reliance on prescripted, linear scenarios. This often results in predictable interactions that fail to fully replicate the dynamic, adaptive, and complex nature of real patient encounters, potentially limiting learner engagement and the ability to respond to individualized student input [16,17]. Consequently, there is a recognized need for more dynamic and adaptable virtual patient models to better prepare learners for clinical practice [18].

GenAI presents a transformative solution to this limitation. A subset of AI powered by LLMs, GenAI can generate context-aware, realistic text, dialog, and even simulate human behaviors in real-time [19,20]. In health care education, integrating GenAI allows virtual patients to move beyond scripts, generating unique, adaptive responses based on learner inquiries and simulating a wider range of patient behaviors and emotional states [7,8]. This capability promises to significantly enhance the authenticity, personalization, and educational value of simulation-based training.

However, the implementation of this novel technology introduces new complexities regarding its design, pedagogical integration, and evaluation. While systematic reviews have established the effectiveness of traditional, scripted virtual patients [21,22] and scoped the broad potential of GenAI in education [5], a critical gap remains. Existing syntheses are unequipped to address the unique research questions (RQs) generated by their convergence. Specifically, the literature lacks evidence-based guidance on how the adaptive, nondeterministic nature of GenAI alters optimal instructional design, what novel evaluation frameworks are required to measure its impact on dynamic clinical reasoning, and how the practical challenges of implementation differ from those of static simulations. This dedicated synthesis is absent. Consequently, without a focused systematic review, evidence regarding the optimal design features, verifiable educational impact, and practical challenges of these advanced tools remains fragmented, hindering evidence-based adoption and focused research development [23,24].

Given these factors, it is essential to conduct a systematic review that not only summarizes the characteristics of GenAI-supported virtual patients but also provides comprehensive guidance for future research in this emerging field. Such a review will help educators, researchers, and policymakers understand the potential benefits and limitations of GenAI-supported virtual patients, thereby facilitating their integration into health care education curricula.

The primary objective of this systematic review is to synthesize the current empirical evidence on the design, implementation, and educational impact of GenAI-supported virtual patients in health care education. Specifically, the review addresses the following RQs:

RQ1. What are the design choices, technological architecture, and educational strategies embedded within GenAI-supported virtual patients?

RQ2. What are the evaluation and educational impact, including benefits, outcomes, and related limitations of GenAI-supported virtual patients?

This review was conducted and reported in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 statement and its expanded checklist (Checklist 1) [25], as well as synthesis without meta-analysis (Multimedia Appendix 1) [26]. The protocol was registered and published in the Open Science Framework repository [27].

The following inclusion criteria were used: (1) original studies published in peer-reviewed journals, (2) focused on evaluating how GenAI-supported virtual patients affect education and training in health care–related disciplines, (3) included measurements of user experience or user outcomes, (4) conducted using quantitative or mixed methods research design, and (5) written in English. The exclusion criteria were the following: (1) book chapters, editorials, short communications, letters, and review literature; and (2) studies focused solely on the technical development of virtual patients without educational evaluation.

A systematic search was performed across 5 databases from their inception to March 19, 2026: CINAHL (via EBSCOhost), MEDLINE (via EBSCOhost), Embase (via Elsevier), Scopus (via Elsevier), and Web of Science Core Collection (via Clarivate). These databases were chosen for their comprehensive coverage of biomedical, nursing, allied health, and interdisciplinary literature. Furthermore, the reference lists of all included full-text papers were manually reviewed to identify additional relevant studies.

The search strategy was developed iteratively for each database to ensure comprehensive coverage of the core concepts of “virtual patient” and “generative artificial intelligence.” For each database, the strategy combined controlled vocabulary (where available, eg, MeSH [Medical Subject Headings] in MEDLINE and CINAHL Subject Headings) with extensive free-text terms, using appropriate field-specific syntax (eg, MH and XB in CINAHL and MEDLINE, :ti,ab in Embase, TITLE-ABS-KEY in Scopus, and TS= in Web of Science). Synonyms for virtual patients and GenAI models were grouped logically, and all free-text terms were searched in the title, abstract, and keyword fields. Searches were executed using the advanced search interface of each platform. The search was limited to records published in English and, where available, to peer-reviewed papers. The search was updated and rerun on March 19, 2026, after incorporating expanded terms and controlled vocabulary as suggested during peer review. The full search strategy is reported in Multimedia Appendix 2 and in accordance with the PRISMA-S (Preferred Reporting Items for Systematic Reviews and Meta-Analyses Literature Search Extension) checklist [28].

Several elements of the PRISMA-S checklist [28] did not apply to our methodology: we did not search databases simultaneously on a single platform (item 2); we did not search study registries (item 3); we did not browse online or print sources (item 4); we did not perform citation searching beyond checking reference lists of included studies (item 5); we did not contact authors, experts, or manufacturers to identify additional studies (item 6); we did not use any other information sources or methods beyond those described (item 7); we did not use published search filters (item 10); we did not adapt search strategies from prior reviews (item 11); we did not set up email alerts or automated updates, but we did manually rerun the search after refining the strategy (item 12); and the search was not formally peer reviewed (item 14). These items are explicitly noted as not applicable in this paper.

All records identified from the database searches were imported into Zotero reference management software for deduplication. The selection process was conducted in 2 phases. First, 2 reviewers (JJ and MZY) independently screened the titles and abstracts of all records against the eligibility criteria. Second, the full texts of potentially eligible studies were retrieved and independently assessed for inclusion by the same 2 reviewers. Any disagreements at either stage were resolved through discussion between the reviewers until consensus was reached. The interrater reliability was calculated using Cohen κ, resulting in κ=0.7, which indicates substantial agreement [29].

A standardized data extraction form was developed in Google Sheets (Google LLC). Further, 2 reviewers (JJ and MZY) independently extracted data from each included study. The extracted data were then cross-checked, and any discrepancies were resolved through discussion. The interrater reliability was κ=0.6, which indicates moderate agreement [29].

Google Sheets (Google LLC) were used for data extraction. The following information was extracted from each of the papers that met the inclusion criteria: (1) publication characteristics: authors, publication year, and source; (2) study characteristics: study design and sample size; (3) intervention characteristics: description of the GenAI-supported virtual patient, including input or output modalities, use of an avatar, duration of interaction, technological details (eg, GenAI model and prompt engineering), and integration with educational strategies or theories; (4) outcomes: all reported outcome measures, including primary and secondary outcomes related to user perceptions (eg, usability, satisfaction, and perceived learning), skills (eg, communication and clinical reasoning), and performance; and (5) key results: main quantitative and qualitative findings as reported by this study’s authors.

To critically appraise the methodological quality and risk of bias of the included studies, a formal assessment was conducted in accordance with PRISMA guidelines (item 11). The JBI (Joanna Briggs Institute) critical appraisal checklists, appropriate to each study design, were used as standardized tools [30]. Specifically, the JBI Checklist for Randomized Controlled Trials was used for randomized controlled trials (RCTs) [31], the JBI Checklist for Quasi-Experimental Studies was used for nonrandomized comparative studies [32], and the JBI Checklist for Analytical Cross-Sectional Studies was used for cross-sectional evaluations [32]. Further, 2 reviewers independently assessed each study. Any discrepancies in appraisal judgments were resolved through discussion to reach consensus. The interrater reliability was κ=0.8, which indicates substantial agreement [29]. The results of this assessment are synthesized narratively and were considered when interpreting the overall strength and validity of the evidence presented in this review.

Current review synthesized findings narratively, thus no common effect measures were pooled across studies due to significant heterogeneity in study designs, interventions, and outcome measures.

A meta-analysis was not feasible due to the limited number of studies and substantial heterogeneity in interventions, populations, and outcome measures. Therefore, a narrative synthesis was conducted. The extracted data were summarized and organized to address the review’s RQs. Findings were structured thematically to describe: (1) the implementation characteristics (design, technology, and educational strategies) of GenAI-supported virtual patients, and (2) their evaluation and educational impact (benefits, outcomes, methodological approaches, and limitations). The synthesis also integrates a discussion of the methodological quality and risk of bias of the included studies.

No formal statistical assessment of publication bias (eg, funnel plot) was performed due to the narrative synthesis approach and the small number of included studies.

A formal assessment of the certainty of the body of evidence (eg, GRADE [Grading of Recommendations, Assessment, Development, and Evaluation]) was not conducted for this narrative review, as its primary aim was to map and characterize an emerging field rather than to estimate a pooled treatment effect.

The initial literature search yielded 2860 studies. After screening the abstracts, 107 papers were selected for further evaluation. Further, 2 authors screened full texts independently, and the screening was then discussed as a group to ensure consensus and make the final selection decision. Ultimately, 15 papers were included in the final analysis after reading the full-text. Figure 1 provides a detailed overview of the systematic search procedure.

All 15 studies included in this review were published in peer-reviewed journals. Research in this area was prominently featured in the JMIR portfolio, which published 6 of the included papers: 3 papers in JMIR Medical Education [33-35], 2 papers in the Journal of Medical Internet Research [36,37], and 1 paper in JMIR Formative Research [38]. Temporally, 8 studies were published in 2024 [33-35,38-42], marking a peak of initial investigative activity, followed by 7 more studies in 2025 and 2026 [36,37,43-47], indicating sustained scholarly interest. The included studies used a diverse range of methodologies. The 2024 cohort predominantly featured cross-sectional (n=5) [33,34,38,39,42] and quasi-experimental (n=2) [35,40] designs, alongside 1 RCT [41]. The 2025‐2026 publications demonstrated a continued diversification of methods, including 2 additional RCTs [45,47], 4 quasi-experimental studies [36,43,44,46], and 1 proof-of-concept observational study [37]. Participant sample sizes varied widely, from smaller feasibility studies (eg, n=6 [40]) to larger-scale evaluations (eg, n=145 [35,43]), and a paired crossover study with 20 medical students [47]. The duration and frequency of interventions with the GenAI-powered virtual patients also differed, ranging from brief, single interactions of approximately 6‐10 minutes [38] to more extended or repeated practice sessions over several weeks [43]. This heterogeneity in publication venues, design, scale, and intervention format reflects the exploratory and rapidly evolving nature of research in this domain.

Moreover, among the 15 included studies, 13 of them explicitly stated obtaining ethics approval from an institutional review board and described informed consent procedures [33-39,41,43-45]. Specific data privacy or security measures were reported less consistently, detailed in only 8 (54%) studies [33-35,37-39,43,46]. Notably, 2 studies [40,42] documented that formal ethics approval was not required for their projects, citing their design as a quality assurance activity and a survey of professionals, respectively.

The methodological quality of the 15 included studies, as assessed by the JBI critical appraisal checklists, demonstrated a spectrum of risk of bias (Table 1). Further, 3 RCTs were judged to have a low to low-moderate risk of bias [41,45,47], benefiting from strong methodologies including randomization, allocation concealment, and blinded outcome assessment. However, the nature of the interventions precluded participant blinding, a common limitation in educational technology studies. In contrast, the body of quasi-experimental [35,36,40,43,44,46] and cross-sectional studies [33,34,37-39,42] presented a higher risk of bias. Common methodological weaknesses across these designs included the use of small, nonrepresentative samples, a lack of control groups, the absence of preintervention or longitudinal outcome measurements, and frequent reliance on unvalidated or self-reported outcome measures. Consequently, while the evidence from RCTs provides a stronger foundation for causal inference regarding the impact of GenAI-supported virtual patients, the overall findings of this review—particularly those related to user perceptions and feasibility—are drawn from a predominantly moderate-risk evidence base. This necessitates a cautious interpretation of the results, as positive outcomes may be influenced by study design limitations and enthusiasm for a novel technology.

To address the RQ1 concerning the design and implementation characteristics of GenAI-supported virtual patients, the following sections analyze the extracted data, which are comprehensively tabulated in Table 2.

The designs of GenAI-supported virtual patients can be classified into three distinct categories: (1) input, (2) output, and (3) avatar. First, in terms of input methods, 11 of 15 studies reported that participants interacted with the virtual patient by entering text [33-35,37,39-41,43,47]. Meanwhile, 5 studies allowed participants to communicate verbally using voice input [36,38,44-46]. Additionally, 1 study incorporated a hybrid approach, enabling participants to use either speech-to-text functionality via a microphone or direct text input through a designated field [42]. Next, the output modalities of the virtual patient corresponded closely to the input mechanisms. In 9 studies, the virtual patient responded to participants via text-based communication [33-35,37,39-41,43,47]. In contrast, 5 studies featured virtual patients capable of generating human-like voice responses [36,38,44-46]. Notably, 1 study highlighted a more versatile approach, where the virtual patient was designed to provide responses to both text and synthesized speech [42].

Regarding the avatar design of the virtual patient in the included studies, 6 studies used a 3D-embodied virtual patient to enhance realism and immersion for participants [36,38,42,44-46]. For instance, a study [38] integrated a mixed reality tool, allowing participants not only to visually perceive the virtual patient within a digital environment but also to physically interact with a corresponding manikin. This setup enabled the virtual patient to display various injuries, movements, and facial expressions aligned with speech production, respiration patterns, and pain-related vocalizations. Similarly, research by Borg et al [36] implemented a virtual patient embedded within a robotic system. Their robot featured a 3-degree-of-freedom neck and an animated face, facilitating flexible head movements and expressive emotional displays. Mool et al [46] also used a 3D avatar, though it was noted to be largely static within the examination room environment. In contrast, the remaining 9 studies used virtual patients without avatars, relying solely on alternative interaction modalities [33-37,39-41,43,47].

In the 15 included studies, interventions involving participants interacting with GenAI-supported virtual patients lasted less than 10 minutes in 5 cases [33,38-41]. Three studies reported intervention durations exceeding 20 minutes [43-45], while the remaining 5 did not specify the duration [34-37,42]. Moreover, 7 studies featured only a single session, regardless of the intervention length [33,35,38,39,42,44,45].

A total of 13 studies explicitly stated that they used an OpenAI GPT model for generating the virtual patient [33-44,46]. Further, 1 study used a fine-tuned model from a different provider [45]. Yu et al [47] evaluated a broader range of models for their AIPatient system, including Claude models (Haiku, Sonnet variants), GPT-family models (GPT-4 Turbo, GPT-4o, and GPT-3.5 Turbo), and open-source models (DeepSeekV3 671B, Qwen3-32B, and LLaMa-3 70B). All studies specified the foundational GenAI model used. A total of 14 studies provided detailed patient case information as part of the prompt to enhance the virtual patient’s responses [33-35,38,40-47]. For instance, in a study [36], the prompt consisted of a structured patient case description, the previous 10 turns of dialog, and an instruction to generate the next line of conversation. Yu et al [47] took a sophisticated approach, integrating real electronic health record data from the MIMIC-III (Medical Information Mart for Intensive Care III) database and incorporating personality profiles based on the Big Five framework to simulate diverse and realistic patient behaviors. In contrast, 1 study used a simpler prompt, wherein the GenAI was instructed merely to assume the role of a virtual patient with a specified condition (eg, respiratory distress) and engage in dialog with participants acting as nurses, without requiring additional case-specific details [39].

A total of 3 studies explicitly referenced established educational frameworks to inform their design. The study by Kim et al [44] applied an instructional design model for AI education and the technology acceptance model. The study by Luo et al [45] used Kolb’s experiential learning cycle and the Calgary-Cambridge communication framework, while the study by Chinwong et al [43] was grounded in the 5As framework for smoking cessation counseling. The study by Kim et al [44] emphasized the critical role of patient case design, noting that such cases serve as the foundational structure of the curriculum, functioning as learning triggers and providing a platform for students to engage in cognitive processes reflective of physicians’ workplace reasoning. Given the high-fidelity patient simulation and the level of control afforded by GenAI, this innovative approach has the potential to enhance medical education curricula, offering valuable benefits for both students and educators. In contrast, the remaining 12 studies did not specify the application of any educational theory to inform the design of the virtual patient or the overall study methodology [33-42,46,47].

To address the RQ2 concerning the educational effectiveness and learner outcomes of GenAI-supported virtual patients, the following sections analyze the extracted data, which are comprehensively tabulated in Table 3.

The current systematic review provides a comprehensive synthesis of the emerging evidence on GenAI-supported virtual patients in health care education, focusing on implementation characteristics, educational impact, and methodological considerations. Building on the descriptive findings presented in the Results section, the following sections offer analytical comparisons across study designs, AI modalities, and educational purposes, and map identified limitations to established learning theories to guide future development.

When examining the evidence base by study design, a clear pattern emerges regarding the strength of causal inferences that can be drawn. The 3 RCTs provide the most robust evidence, demonstrating significant improvements in clinical decision-making [41], ophthalmology history-taking skills [45], and medical history-taking performance [47] attributable to GenAI-supported virtual patient interventions. In contrast, the cross-sectional and quasi-experimental studies, while consistently reporting positive user perceptions and self-reported skill gains [33,34,38,39,42,46], are more susceptible to enthusiasm bias and cannot establish definitive causal links between the intervention and learning outcomes. This methodological gradient underscores the need for more rigorous, controlled trials to move the field beyond proof-of-concept toward evidence-based practice.

Critically, the overall evidence base is characterized by small sample sizes, brief intervention durations (often single sessions under 10 minutes), and a near-complete absence of longitudinal follow-up. Consequently, while the studies demonstrate that GenAI-supported virtual patients are feasible and acceptable to learners, claims regarding their educational effectiveness must remain preliminary. The current findings support feasibility and acceptability more strongly than demonstrable learning gains or transfer to clinical practice.

Analysis by AI modality reveals differential alignment with educational objectives. Studies using embodied, voice-based virtual patients [36,38,44-46] predominantly targeted communication skills and emotional realism as primary outcomes. For example, Borg et al [36] found that a social robotic platform with 3D embodiment was rated significantly higher for authenticity and emotional skill training compared to a conventional computer-based platform, suggesting that physical presence and nonverbal cues may be particularly valuable for interpersonal skill development. In contrast, text-based virtual patient studies [33-35,37,39-41,43,47] more frequently targeted clinical reasoning, diagnostic accuracy, and history-taking performance. Yu et al [47] directly compared their text-based AI patient to human-simulated patients and found comparable or superior support for clinical reasoning, indicating that for cognitive skills such as information gathering and diagnostic thinking, sophisticated text-based interactions may be as effective as, or more effective than, human simulations. This modality-outcome alignment has practical implications for instructional design: educators should select or design GenAI virtual patient platforms based on the specific competencies they aim to develop.

When considering educational purpose, studies targeting communication skills [38,40,44,46] consistently reported improvements in learner empathy, interviewing technique, and patient interaction quality. For instance, Maquilón et al [38] found that medical first responders rated the virtual patient’s voice interactions as natural for assessing physiological state, while Mool et al [46] observed that some students anthropomorphized the avatar, using the patient’s name and inferring attitudes—behaviors indicative of authentic communication practice. Studies focused on clinical reasoning [36,41,47] demonstrated gains in diagnostic accuracy, information gathering, and decision-making, with Brügge et al [41] showing that feedback-enhanced interactions led to significantly higher clinical reasoning scores. This pattern suggests that GenAI-supported virtual patients can be effectively tailored to specific educational objectives, with modality and design features aligning with targeted learning outcomes.

From the perspective of experiential learning theory [48], the lack of longitudinal consistency and evolving patient states prevents learners from engaging in the full cycle of concrete experience, reflective observation, abstract conceptualization, and active experimentation across repeated encounters. Without memory across sessions, students cannot build upon prior interactions or observe the consequences of their clinical decisions over time—a core component of developing clinical expertise. This limitation suggests that future GenAI virtual patient designs should incorporate persistent patient states and longitudinal case progression to support complete experiential learning cycles.

Cognitive load theory [49,50] provides another lens for interpreting current limitations. The technical glitches, inconsistent responses, and need to troubleshoot AI interactions observed in several studies [38,43,46] impose extraneous cognitive load, diverting mental resources away from the germane load essential for schema construction and clinical reasoning development. Reducing technical unpredictability through more robust prompt engineering and model selection, as demonstrated by Yu et al [47], could minimize extraneous load and optimize cognitive resources for learning.

Furthermore, the absence of adaptive emotional and behavioral complexity limits opportunities for the sustained, feedback-driven practice necessary to refine complex interpersonal skills. Well-established principles of skill acquisition emphasize the importance of repeated engagement with authentic tasks, immediate feedback, and progressive challenge—elements that current GenAI virtual patients only partially provide. Incorporating dynamically adjusting emotional states based on learner interactions, informed by frameworks such as the Big Five personality model used by Yu et al [47], could better support the deliberate practice of communication and empathy.

The implementation of GenAI-supported virtual patients in health care education carries significant practical implications. Educators and curriculum designers must carefully consider the technological infrastructure required to support diverse interaction modalities—text, voice, and hybrid formats—to cater to varied learning preferences and replicate realistic clinical scenarios. The integration of dynamic, AI-generated responses demands ongoing technical oversight and periodic updates to ensure the content remains accurate and relevant. Furthermore, the relatively brief interaction sessions observed in this study suggest that more extended, immersive simulations are needed to mirror the complexity of real-world clinical practice. This calls for interdisciplinary collaborations between educators, clinicians, and technology developers to ensure that practical implementations not only leverage the technical advantages of GenAI but also align with educational objectives and clinical standards.

The finding that only 3 of the 15 included studies were explicitly grounded in established educational frameworks highlights an important area for future research. The limited adoption of frameworks indicates that many current approaches are primarily driven by technological innovation rather than pedagogical insight. By integrating robust theoretical perspectives, the design of GenAI-supported virtual patients can be enhanced to facilitate richer, more targeted learning experiences that promote the development of clinical reasoning and decision-making skills. Furthermore, theoretical models can help in formulating clear hypotheses about how these advanced simulations influence learner outcomes, guiding both the design of interventions and the evaluation of their effectiveness [51]. Encouraging further interdisciplinary studies that explicitly link technology-driven interventions with established learning theories will be essential in building a more comprehensive understanding of how these innovations can transform health care education.

For research design, only 5 of the 15 studies incorporated experimental designs with control groups, allowing for direct comparisons between traditional simulation methods and novel AI-driven interventions [35,36,41,45,47]. This experimental rigor has been crucial in establishing a baseline understanding of the potential advantages of GenAI-enhanced simulations, especially in relation to improving communication skills and clinical reasoning. The remaining 10 studies typically used within-group designs with mixed methods or solely quantitative approaches, providing insights into user perceptions and immediate learning outcomes [33,34,37-40,42-44,46]. As highlighted in the synthesis above, the current evidence base supports feasibility and acceptability more strongly than conclusive educational effectiveness. Future research must therefore prioritize larger, rigorous trials with long-term follow-up and objective measures of skill transfer to clinical practice.

To provide a more balanced perspective, the potential benefits of GenAI-supported virtual patients must be considered alongside their well-documented risks and ethical challenges. Our findings regarding positive learner perceptions and skill development exist within a context of significant technological limitations. Key concerns include algorithmic bias inherent in the training data of LLMs, which could reinforce stereotypes or lead to culturally insensitive patient portrayals, thereby misinforming clinical empathy and communication [52,53] Data privacy and security present another critical challenge, as sensitive health information used in prompt engineering or generated during simulated dialogs requires robust safeguards to comply with regulations such as HIPAA (Health Insurance Portability and Accountability Act) or GDPR (General Data Protection Regulation) [54,55]. Most critically, the tendency of GenAI models to generate plausible but incorrect or fabricated information—known as “hallucinations”—poses a direct threat to clinical accuracy [56]. In a health care education setting, an AI virtual patient providing factually wrong symptoms, pathophysiology, or treatment responses could seriously compromise foundational medical knowledge and patient safety. Therefore, the implementation of these tools necessitates rigorous validation frameworks, ongoing human oversight, and clear institutional guidelines to mitigate these risks, ensuring that innovation in simulation does not come at the cost of pedagogical integrity or ethical responsibility.

The risk of bias assessment conducted for this review underscores a critical consideration when interpreting its findings. While 3 RCTs demonstrated relatively strong methodological rigor (low to low-moderate risk) [57], most of the evidence derives from quasi-experimental and cross-sectional studies with moderate to high risk of bias [58]. This methodological landscape indicates that the current evidence base is still in a formative, proof-of-concept stage [59]. The prevalent limitations—such as small sample sizes, absence of control groups, lack of blinding, and reliance on unvalidated or self-reported outcomes—suggest that the positive results regarding user acceptance, perceived learning, and skill improvement should be viewed as promising preliminary signals rather than conclusive evidence of efficacy [60-63]. These design weaknesses increase the risk of overestimating positive effects due to confounding, measurement bias, or participant enthusiasm for novel technology [64,65]. Therefore, the synthesized findings, particularly those related to educational impact, must be interpreted with appropriate caution. This appraisal directly informs the primary recommendation of this review: future research must prioritize methodological robustness, including larger-scale randomized designs with active control groups [66], longitudinal follow-up [67], and the use of validated, objective outcome measures [65] to establish a more definitive evidence base for the educational effectiveness of GenAI-supported virtual patients.

Furthermore, our analysis identified variable reporting of ethical considerations—such as institutional review board approval and data security measures—across the included studies. While systematic reviews themselves do not require ethical approval, transparency in primary research is a cornerstone of methodological rigor and trustworthiness [68]. Moving forward, consistent and explicit ethical reporting should be considered a standard in this domain, especially when research involves simulated patient interactions and learner data, to ensure the credibility and safe translation of findings into educational practice [53,54].

This review has several limitations that should be addressed when interpreting its findings. First, while our systematic search was comprehensive across several databases, the inclusion criteria limited the review to English-language studies, potentially excluding relevant research published in other languages or in alternative repositories. Second, many of the primary studies included in this review are constrained by methodological limitations, such as small sample sizes, short simulation durations, and an overreliance on self-reported outcomes, which restrict the generalizability of the findings. Third, the heterogeneity in study designs and assessment tools across the interventions complicates direct comparisons and synthesis of outcomes. Furthermore, our search was limited to published, peer-reviewed literature and did not include gray literature sources such as clinical trial registries or preprint servers. While this decision aligns with standard systematic review practices, it is particularly consequential in a rapidly evolving field where early innovations often first appear as preprints or technical reports. Consequently, the review may not capture the most recent developments or emerging design approaches that have not yet undergone formal peer review. Future updates to this review should consider expanding the search to include gray literature as the evidence base matures. Lastly, the current review only included 15 papers; this limited number of studies may not capture the full spectrum of innovative practices and outcomes in this rapidly evolving field, thereby constraining the robustness of the conclusions drawn.

In summary, this review confirms that GenAI-supported virtual patients offer notable advances in adaptability and interactivity. This study is innovative as it constitutes the first dedicated synthesis of this specific technological application in health care education. It differs from prior reviews of virtual patients or GenAI by focusing exclusively on their intersection, thereby isolating the unique capabilities and challenges introduced by GenAI. The review brings to the field a foundational framework that classifies key design dimensions, evaluates educational impact, and identifies critical gaps, setting a clear agenda for subsequent research. The real-world implications are significant: for educators and technologists, it provides an evidence-based roadmap for developing more effective, theory-informed simulations; for institutions, it highlights the practical considerations and potential transformative value of integrating these tools to modernize clinical skills training and address scalability in health care education.