To guide the methodological structure of this review, we used two widely recognized frameworks: PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) and PICO (Population, Intervention, Comparison, Outcome) (31). PRISMA was selected to promote clarity in literature selection and transparency in reporting (32), structuring the study identification and selection into four stages with specific steps at each phase.

The PICO methodology was chosen because it allows for the formulation of structured inclusion criteria and helps maintain thematic consistency among selected studies (33). It is particularly useful for identifying articles focused on interventions that align with the scope of the review. This PICO framing, including the research question, was defined a priori before data extraction, and guided both the construction of the search string and the subsequent screening process. The population considered included learners across the continuum of medical and health-professions education, namely undergraduate medical students, residents, and fellows in graduate medical education, and practicing clinicians engaged in continuing professional development. No limits were imposed on clinical specialty, provided that participants were enrolled in a structured educational or training activity.

Interventions were grouped into three categories: (i) AI-only educational applications (e.g., natural language processing tools, generative models, intelligent tutoring systems); (ii) XR-only applications (virtual, augmented, or mixed reality–based simulations); and (iii) integrated AI–XR interventions in which both components were used within the same educational activity. Comparators included traditional teaching methods and non-intervention conditions, and outcomes focused on learning effectiveness. Four outcome domains were defined a priori: knowledge (scores on written or structured theoretical tests), skills (performance on practical or procedural tasks), performance (objective indicators such as error rates, completion time, or simulator-derived scores), and engagement (self-reported or behavioural indicators of motivation and participation). The main assessment instruments for each study are summarized in Table 5, and the overall PICO framing in Table 1.

Using the PICO model, the research question was iteratively refined to align with the objectives of this review: to explore the educational impact of Artificial Intelligence and Extended Reality technologies in medical and healthcare training. The final research question guiding this review was:

“In empirical studies published between 2019 and 2024, how have AI-only, XR-only, and combined AI–XR technologies been used in medical and health-professions training, what learning outcomes have been reported, and what implementation challenges have been described, with specific attention to the limited number of fully integrated AI–XR interventions?”

Given the limited number of fully integrated AI–XR studies, we also retained AI-only and XR-only interventions to contextualize the emerging evidence base and identify design patterns for future AI–XR ecosystem development. Guided by our research question and the PICO framework, we implemented a structured Boolean search organized into three concept clusters combined with AND—(1) AI-related terms (e.g., artificial intelligence, natural language processing, generative AI), (2) medical and health-professions education and training, and (3) instructional modality (e-learning, simulation, virtual reality)—with synonyms within each cluster combined with OR. The complete, database-specific search strings are summarized in Table 2.

To ensure a comprehensive and multidisciplinary collection of articles, three major academic databases were selected for this review: SCOPUS, PubMed, and IEEE Xplore. Each database contributes uniquely to the intersection of healthcare, education, and emerging technologies, making them essential for capturing diverse research perspectives.

SCOPUS was selected for its extensive indexing of peer-reviewed journals in health sciences, education, and engineering.

The search strategy employed in these databases yielded a diverse set of studies spanning the technological, educational, and clinical dimensions of AI in medical education. Boolean queries were applied uniformly, and the results reflect the multidisciplinary scope of the field. Table 2 presents a concise summary of the query structure and the number of articles retrieved from each source, providing a fundamental dataset for the screening and eligibility stages detailed in the following sections.

The volume and temporal evolution of these retrieved records provide insight into how interest in AI within medical education has developed over time. As shown in Figure 1, there is a clear increase in the number of records identified in all three databases from 2020 onwards, with SCOPUS consistently reporting the highest annual counts, followed by PubMed and IEEE Xplore. This trend reflects broader growth in research on AI and immersive digital tools in medical education rather than the characteristics of the final 13 studies included in the review.

The identification phase involved a systematic search of the three main databases mentioned above: SCOPUS (n = 258), PubMed (n = 84), and IEEE Xplore (n = 50). The search focused on literature specifically examining how AI technologies are integrated into teaching, learning, and assessment processes within medical education.

Figure 2 shows how the initial search yielded 392 records. The search strategy was intentionally designed to be highly sensitive, in line with PRISMA-oriented guidance, to minimize the risk of missing empirical AI- and XR-based educational interventions. Consequently, many records identified at the identification stage were concept papers, narrative or systematic reviews, or clinical AI studies without an educational component, which were subsequently excluded during the eligibility assessment.

After removing duplicates, 336 records remained. During title/abstract screening, records lacking essential metadata (e.g., abstracts or keywords) were excluded (n = 146), leaving 190 records for full-text review. At full-text screening, non-peer-reviewed formats (books, book chapters, conference papers, letters, and editorials) were excluded (n = 93), yielding 97 eligible articles. Finally, we excluded 3 studies not related to AI and Medicine, 67 not related to Education/Training, and 14 review articles, resulting in 13 studies included in the intervention analysis.

After removing duplicates, 336 records were available for the initial assessment. The goal of the screening phase (title/abstract screening) was to retain only records clearly relevant to the application of AI in medical and health-professions education for full-text eligibility. Each record was checked for title, abstract, indexed keywords, and basic bibliographic completeness.

Records lacking essential metadata needed to judge relevance—principally abstracts and/or indexed keywords (a frequent issue in technical databases such as IEEE Xplore)—were excluded.

All exclusions were logged to ensure transparency and reproducibility. In total, 146 records were excluded at this stage, leaving 190 records for full-text eligibility assessment. The outcomes of the screening step are summarized in Figure 2.

The eligibility process involved the exclusion of 93 records based on publication type and source format. The excluded articles included books, book chapters, conference proceedings, editorials, letters, and commentaries, which, during the full-text review, were identified as not contributing information to the inclusion criteria. Recognizing that some of these sources provided important conceptual descriptions or theoretical perspectives on AI and its potential applications in medical education, they were rejected due to their preliminary nature, which often limits methodological detail and omits follow-up results or student assessment data.

In the case of editorials and letters, while sometimes relevant to ongoing debates in the field, they were also excluded due to their opinion-based structure and lack of data-driven findings. Books and book chapters, although often comprehensive, present challenges related to accessibility, inconsistent peer review standards, and variability in how interventions and outcomes are reported.

The exclusion of these 93 articles was necessary to preserve the analytical integrity and focus of the review, maintaining the focus on information extraction specifically in original articles. Therefore, only full-text empirical studies were eligible for further analysis. After this filtering process, 97 full-text articles were retained, which constitute the eligibility stage, where their relevance to AI applications in medical education and the presence of measurable learning outcomes were further assessed.

Of the 97 full-text articles assessed for eligibility, 13 studies were included in the final analysis. The selection of this resulting number of studies for inclusion was the result of a rigorous three-stage selection process, designed to isolate only those studies that met all thematic, methodological, and empirical criteria related to the central focus of the review: the use of AI in medical education.

The first stage of exclusion focused on articles unrelated to AI and medicine, of which 3 were identified, leaving 94. The eliminated studies were related to AI in a general or metaphorical sense, but did not involve actual AI systems, models, or algorithms in medical education contexts and focused on training in disciplines not directly related to medicine.

In Stage 2 of the eligibility assessment, we excluded technical and clinical journal articles that addressed the development or validation of AI systems for diagnosis, risk prediction, or clinical workflow optimization, but that did not involve students or evaluate educational interventions and therefore did not align with the review objective of synthesizing evidence on AI-based teaching and learning practices. These records were coded as having no AI-based educational applications because they did not report implemented AI, XR, or combined AI–XR tools used in teaching, learning, assessment, or simulation for health-professions learner. Thirty-four articles were excluded at this stage, leaving 67 studies for the final filter.

To maintain the focus on original and empirical research, review articles, both narrative and systematic, were then excluded from the dataset. Systematic, scoping, and narrative reviews were therefore treated as non-eligible study designs and were used only for contextual background and to cross-check whether our search had missed primary studies (n = 14 excluded). While informative, these articles summarized existing literature rather than presenting new data or intervention results. The exclusion of these studies means that the final synthesis is based solely on primary studies with defined populations, interventions, and measurable learning outcomes.

The 13 articles ultimately selected constitute the inclusion stage. These are peer-reviewed, intervention-based research on the applications of AI in medical education, which included a quantitative or qualitative assessment of the educational impact.

In terms of study design, the presence of a control group was not used as an inclusion criterion. Many AI- and XR-based educational interventions are still reported as pilot or feasibility work and therefore lack formal comparators. During data extraction, the presence and type of comparator or control condition were recorded for each study and, when applicable, classified into one of four categories: (i) no formal comparator (single-group or pre–post design), (ii) traditional instruction or standard curriculum, (iii) alternative technology-enhanced instruction without AI or immersive XR, and (iv) historical or convenience cohorts.

All included studies underwent a structured quality appraisal to assess the robustness of their design and reporting. Each article was evaluated using design-appropriate critical appraisal checklists (e.g., tools for randomized and quasi-experimental studies, observational designs, and qualitative research). When assessments differed, they were reconciled through discussion until consensus was reached. The appraisal focused on key domains such as study design and setting, sampling strategy and sample size justification, clarity and validity of outcome measures, handling of confounding factors, and transparency of data analysis and reporting. Given the heterogeneity of study designs and outcomes, we did not compute a pooled quality score. Instead, studies were categorized qualitatively as offering lower, moderate, or higher methodological robustness, and these judgements informed the synthesis presented in the “Methodological Considerations and Quality of Evidence” section.

Table 3 presents an overview of the studies, their variables, methodologies, and sample sizes. Five studies focused on engagement, defined as the emotional and cognitive involvement of learners in learning activities. These interventions often used immersive or personalized environments to increase motivation and attention (34) surveyed 406 healthcare professionals and found widespread enthusiasm for XR-based training environments. However, the cross-sectional design limited causal inference. Similarly (38), explored how AI and VR were perceived in virtual interviews, with results indicating adaptability and a positive reception among 112 participants. In the study by (45), they demonstrated the reach of an international e-learning platform with NLP/ML support, with over 1,600 participants. Despite the large sample size, the study acknowledged limitations in assessing long-term impact (46). analyzed radiology training during the COVID-19 pandemic, highlighting the role of virtual platforms in maintaining student engagement (36), while focusing on communication skills, reported that AI-generated case simulations increased student participation due to their interactivity and realism.

Figure 3 details that engagement-oriented studies represented 38% of the sample. These findings suggest that AI and XR can effectively foster student participation, especially when traditional formats are disrupted or inaccessible. Skills development emerged as the most frequently assessed educational variable, appearing in seven of the 13 studies (54%) (35). conducted a pilot study on the use of ChatGPT to simulate conversations between radiology technicians and claustrophobic patients. Although the sample size was limited (n = 6), the study demonstrated the potential for developing communication skills through AI-generated dialogue (36). These studies used standardized AI-generated cases in veterinary training (n = 237) and reported improvements in confidence and fluency during client interactions.

Other studies, such as (40), employed a virtual reality simulation program for pediatric residents, combining theory with practice. In this study, participants showed improved task performance, although the institutional scope was limited. In another application (41), offered a four-week AI course that integrated clinical and evidence-based medicine content, fostering critical thinking despite small group sizes (43). Their research encouraged students to design their own clinical decision support systems, integrating AI tools into the curriculum design: a promising approach to skills development through hands-on experience (44). They developed the VIOLA platform for diagnosing diabetic retinopathy, training 150 participants via an online interface. Finally (42), described an AI-powered VR platform, currently under development, and proposed future training in virtual patient encounters.

Only two studies (15%) explicitly measured performance, defined as the demonstration of competence in clinical tasks or simulations (37). used machine learning models to classify students' progress in laparoscopic virtual reality simulations. Despite the innovative analysis, the small sample size (n = 23) and reliance on subjective classification limited the robustness of the conclusions (39). tested Virtual Augmented Surgical Navigation (VASN) in otolaryngology simulation. The intervention improved performance in surgical planning but was limited by the sample size (n = 15) and simulation fidelity. As shown in Figure 4, performance remains the least explored outcome in the current AI/XR education literature.

Two studies were identified as pilot initiatives, representing initial experiments to explore feasibility and usability (35). conducted a pilot study on the use of ChatGPT for radiologist and patient simulations, and (40) implemented a VR pilot program; two other studies were in the development stage, where no outcome assessment had been performed (42). proposed an AI-powered VR platform but did not present student data. The existence of such nascent projects highlights the innovative nature of this research area but also points to the need for more rigorous validation frameworks.

The distribution shown in Figure 4 reflects a steady flow of innovation, progressing gradually from exploratory design to empirical validation. However, the relatively low number of pilot and development projects could imply a publication bias toward studies with proven technologies, which might overlook the value of documenting conceptual innovations or prototypes.

In general, studies consistently documented improvements in educational performance, student confidence, and skill acquisition (34). conducted a cross-sectional survey in Pakistan to assess perceptions and readiness for integrating XR into medical education. The results indicated that 83.8% of participants believed XR could effectively improve the quality of education and patient care, observing statistically significant correlations between familiarity with XR and favorable attitudes toward its use in diagnostic and surgical procedures.

Table 4 summarizes the observed intervention effects together with study design and comparator. In this review, effects were classified as “positive” when the original study reported improvement in at least one predefined learning outcome (knowledge acquisition, skill development, performance, or engagement), either relative to baseline in pre–post designs or compared with a control or alternative condition, using the significance criteria specified by the authors. Outcomes were labeled as “mixed/neutral” when no clear improvement was observed or when findings differed across domains.

Regarding study designs and comparators, we observed a predominance of single-group or pre–post evaluations. Only a minority of studies implemented a formal control group, typically based on traditional lectures or standard skills training, while a smaller subset compared AI or XR interventions against other digital resources or simulation modalities. The type of comparator used in each study is summarized in the “Comparator/control condition” of Table 4.

Similarly (40), in a randomized controlled pilot trial, demonstrated that residents participating in a VR-based behavioral health training module showed statistically significant improvements in motivational interviewing (MI) and anticipatory behavioral health guidance (OAHG) skills. These improvements were quantitatively supported by a higher frequency of open-ended questions and MI-aligned behaviors, illustrating the effectiveness of immersive simulation as a teaching tool.

Another noteworthy contribution comes from (35), who used ChatGPT-3.5 and ChatGPT-4 to simulate interactions with patients in a radiographic environment. This pilot case study showed a 96.7% success rate in managing virtual patients experiencing claustrophobia during magnetic resonance imaging (MRI), with communication styles adapted to the different experience levels of the radiologic technologists. This innovative application of LLM models adds a new dimension to preclinical training by reinforcing communication skills between students and patients.

Figure 5a shows that 11 of the 13 studies (84.6%) reported a positive impact on learning or teaching delivery, whereas 2 (15.4%) were mixed/neutral. Figure 5b shows the varied geographical distribution of the studies, with contributions from North America (n = 4), Europe (n = 4), Asia (n = 1), and one multicenter study. Four studies did not specify their location, which limited a more detailed geographical analysis. The United States had the largest representation (n = 3), with studies focusing on innovative applications such as digital radiology training during COVID-19 (46) and competency-based AI training (41). The multicenter evaluation in (45) revealed the scalability of e-learning platforms and explored the use of natural language processing to analyze user feedback in 24 countries, providing insights into the global applicability of AI-enhanced learning.

Despite the positive results in the vast majority of studies, almost all reported significant barriers hindering smooth adoption. Table 4 shows technological complexity and system integration problems detailed in (39) and (42), as well as concerns about data privacy and a lack of regulatory clarity observed in studies (34) and (36). Another key barrier, according to (41) and (45), is inadequate infrastructure and training for carrying out study activities, compounded, as in the case of (40), by high costs and resource limitations.

A recurring theme in various studies is the ethical and legal ambiguity associated with the use of AI and XR in educational settings. Study (34) identified privacy concerns as one of the main barriers (mentioned by 63.9% of participants), along with technological underdevelopment (70%) and insufficient XR training among healthcare professionals (45.8%) (36)., detailing a centralization in the use of AI in veterinary education, observed that upper-level students showed resistance due to uncertainty and a lack of confidence in AI tools, pointing to a systemic need for early training in AI digital literacy. Similarly (43),, through their elective AI-CDSS course implemented in France, reported logistical barriers to implementation, although student participation remained high.

Table 5 shows that current evidence reveals a dynamic integration of AI and immersive technologies across a wide range of medical education settings; these innovations offer promising strategies for skills acquisition, communication training, and clinical reasoning. As shown in Figure 6, general medical education represented the most common area of ​​focus, encompassing almost half of the reviewed studies, a total of 6. This finding reflects a growing recognition of the critical role that digital health literacy plays in modern clinical practice. Generalist programs, such as those implemented by (34) (41),, and (43), frequently addressed key concepts of AI, clinical decision support systems (CDSS), telemedicine, and data ethics. On the other hand, the study by (41) presented a four-week interactive course that used clinical practice simulations and online modules to teach medical students the application of AI in patient care. This course demonstrated a high level of participation and knowledge acquisition, underscoring the importance of integrating AI literacy from the early stages of medical curricula.

Surgery was explored in two studies (37, 39). In the first study, a VR-based structured laparoscopic curriculum was evaluated using machine learning classifiers to assess student performance. The incorporation of explainable AI (XAI) models into the feedback process allowed for objective, real-time performance monitoring (39). In their study, augmented surgical navigation (VASN) was used to enhance training in functional endoscopic sinus surgery (FESS). Their results detail that participants demonstrated significant improvements in their skill and confidence scores, supporting the effectiveness of AI-assisted simulation environments for surgical training.

Radiology was the focus of two additional studies (35). In their study, generative AI (ChatGPT-3.5 and 4.0) was used to simulate communication between radiology technicians and claustrophobic patients. The study highlighted the usefulness of AI in developing interpersonal skills through natural language interaction, an area often overlooked in traditional radiology training. In the same vein (46), researchers analyzed the impact of COVID-19 on radiology case volume and instructional design, observing a shift toward online remote learning platforms. These studies underscore how AI can facilitate both the technical and interpersonal aspects of radiology training.

Other fields included pediatrics (40), ophthalmology (44), and veterinary medicine (36). For example (40), a randomized controlled trial was conducted that incorporated virtual reality simulations to improve behavioral health anticipatory guidance (BHAG) and motivational interviewing skills among pediatric residents. The results indicated statistically significant improvements in communication skills, illustrating the potential of immersive technologies in behavioral pediatrics. In ophthalmology (44), the study developed a self-learning digital platform for diabetic retinopathy detection. The program offered certification and interactive learning tasks, and more than 150 healthcare professionals completed the course, demonstrating the scalability of this type of model. In contrast, research by (36) presented veterinary students with AI-generated standardized clients, providing valuable information on integrating AI into veterinary communication training. The students responded positively but identified a clear need to increase AI literacy within veterinary education.

The studies analyzed employed diverse training modalities, reflecting both institutional resources and pedagogical preferences. These modalities included self-paced online modules (44), structured simulation labs (39), short workshops (36), randomized controlled trials (40), and e-learning courses with natural language processing (NLP)-based assessments (45). For example (34, 38), were survey-based and did not involve direct educational interventions, but rather assessed perceptions and attitudes toward AI and XR in medical training settings.

The duration of the educational interventions varied significantly, from single-day or single-session labs (36) to modules lasting several weeks (41). In study (40), monthly instructional sessions were conducted over a four-month period, combining theory with repeated virtual reality practice. In comparison with studies (35, 37, 46), these studies did not report the duration of the intervention, highlighting the need for standardized reporting practices in medical education research.

Assessment tools ranged from validated questionnaires and psychometric scales to advanced analytical frameworks, including machine learning. Study (35) evaluated simulation success rates and AI error detection, while study (37) used predictive models to assess skill progression in laparoscopic tasks. Study (41) relied on weekly questionnaires and task ratings to monitor student learning outcomes, and study (40) employed pediatric psychologists to code residents' behavior in simulated encounters. Qualitative feedback was also common (43). incorporated student self-reports and project engagement metrics to measure curriculum effectiveness, while (45) applied sentiment analysis and theme modeling to extract qualitative feedback from large-scale courses.

Across the 13 studies, knowledge acquisition was typically measured through written tests or online quizzes; skill development and performance were captured through OSCE-style checklists, simulator-derived scores, error rates, and task-completion times; and engagement was assessed using Likert-type satisfaction or engagement scales, usage statistics from digital platforms, and qualitative feedback. This mapping between outcome domains and assessment instruments is summarized in Table 5.

Finally, some studies, such as (42), described digital education platforms under development, including “medical tr.AI.ning,” a modular tool designed to simulate first-person clinical decision-making using AI-powered virtual agents. These projects point to the future of medical education: scalable, customizable, and highly interactive environments powered by AI and XR.

Several studies have demonstrated the positive impact of AI and XR technologies on student skills development and engagement (35). They used ChatGPT to simulate interactions between radiologists and patients, improving communication strategies during MRI scans (36). They showed that veterinary students benefited from AI-generated clinical cases and standardized patients, which facilitated communication training through immersive and contextualized experiences.

Other interventions, such as the study (39), integrated VASN augmented surgical navigation into endoscopic sinus surgery training, demonstrating improvements in technical performance and confidence in the procedure (40). They implemented virtual reality simulations to improve motivational interviewing and behavioral health counseling skills in pediatric residents (41). They introduced a structured AI module focused on evidence-based medicine and clinical integration, which improved students' conceptual and applied understanding. Their approach, based on reverse design and modular instruction, illustrates how AI literacy can be taught even to students with no prior technical knowledge (43). They went further by involving students in the design of AI-based clinical decision support systems, fostering critical thinking and digital leadership.

These findings point to several directions for advancing medical education with AI and XR. Interventions should be aligned with existing competency-based frameworks by mapping AI- and XR-enhanced activities to specific knowledge, skills, and professional behaviors, rather than treating them as isolated add-ons. In many of the most promising examples, AI or XR is embedded within longitudinal learning sequences, such as combining introductory modules on AI literacy with progressively more complex simulated cases and opportunities for supervised decision-making. Immersive simulations and AI-driven feedback can be directed toward competencies that are difficult to practice in routine clinical placements, including communication in high-stakes scenarios, management of rare events, and interprofessional collaboration. Several studies also highlight the value of involving students as co-designers of AI tools and learning activities, which may support digital professionalism, critical appraisal of algorithms, and more responsible adoption of these technologies in future clinical practice.

The formal quality appraisal confirmed substantial methodological heterogeneity across the 13 included studies, both in research design and in the assessment tools employed. This diversity provides a broad overview of approaches to AI and XR in medical education but also limits comparability and constrains the strength of inferences that can be drawn from the available evidence.

A large proportion of studies relied on observational or descriptive designs, such as cross-sectional surveys (34, 35, 38) and uncontrolled educational evaluations (43, 44). These designs are useful for exploring initial perceptions, feasibility, and acceptance of new technologies, yet their capacity to establish causal relationships or accurately quantify educational impact is limited. The absence of control groups in many of these investigations, together with a frequent reliance on subjective measures—such as self-administered questionnaires or satisfaction surveys—reduces the robustness of the findings.

A smaller subset of studies adopted more rigorous designs, including randomized controlled trials (40), prospective trials (39), and educational innovations with systematic performance evaluation (41). These approaches offer comparatively stronger evidence and allow more credible inferences about the effects of AI- and XR-based interventions on clinical learning. Nevertheless, even within this group, the quality of the evidence is often constrained by small sample sizes, single-centre settings, and restricted follow-up periods. Long-term assessments of knowledge retention, skills transfer to real clinical environments, and sustained behavioural change were rarely reported, and outcome measures were not always based on validated instruments.

Another methodological consideration is the diversity of technologies and implementations evaluated. Some studies focused on virtual reality simulations (39, 40), others on conversational AI tools (35, 36) or e-learning platforms (44, 45). Across these contexts, limited faculty training, inadequate financial and technical resources, and logistical barriers were frequently highlighted. These constraints not only influence the success of implementation but also impact study design, sample recruitment, and the quality of data collection.

These features indicate that the current evidence base should be interpreted as preliminary and largely hypothesis-generating rather than definitive. While signals of benefit for AI and XR in medical and health-professions education are consistent across studies, the overall certainty of the evidence is low to moderate at best. Stronger, more standardized methodological approaches will be required to determine the true magnitude and generalizability of the educational effects attributed to these technologies.

Table 6, a comparison of this review with existing literature, reveals convergences and differences in thematic focus, methodological orientation, and conceptual contributions to understanding how digital technologies—particularly AI and XR—are transforming medical education. While several previous reviews have addressed the integration of emerging technologies in medical or health-related training, this review synthesizes empirical studies across diverse clinical settings, educational modalities, and geographical contexts, offering a more comprehensive and integrative perspective.

Some reviews have focused on specific tools or technologies, such as the use of ChatGPT in radiology training (47). While that review highlights the potential of generative AI to produce assessments and learning materials and points out risks such as bias and misinformation, it focuses exclusively on content generation rather than broader pedagogical outcomes. In contrast, the present study explores AI in relation to student engagement, clinical skill acquisition, and curriculum design, offering a systemic perspective that goes beyond a single educational tool.

Other studies have emphasized feedback mechanisms in clinical education (48), particularly the role of digital platforms, mobile applications, and virtual reality in facilitating personalized, real-time feedback. While our review acknowledges the value of interactive learning environments, it places greater emphasis on the curricular and infrastructure implications of adopting AI and XR, identifying challenges such as a lack of faculty training, limited technological infrastructure, and ethical considerations that may hinder implementation. Both reviews agree on the need for governance frameworks that ensure data privacy and the effectiveness of the tools, often in specific specialties such as radiology (17), neurosurgery (49), or anesthesiology (51). These reviews typically highlight the pedagogical advantages of XR for procedural training, spatial awareness, and simulation-based learning. Our review incorporates these findings, but expands the analysis across various disciplines, demonstrating that XR not only supports technical skills, but also fosters engagement and interactive learning in diverse clinical contexts, such as pediatrics, ophthalmology, and general medicine.

Broader thematic syntheses have also emerged. One review categorizes technological innovations in diagnosis, treatment, and education (50), including the use of the metaverse and patient-centered design. While this typology is useful for understanding the broader application of digital tools in healthcare, it lacks the curriculum specificity and empirical grounding present in our review. Similarly, reviews addressing the metaverse and telemedicine (54), or transformations during the COVID-19 pandemic (55), provide contextual information but offer little depth regarding educational implementation or learning outcomes.

A recurring theme in other reviews is the need to balance technological advances with traditional training approaches. This is particularly evident in areas such as regional anesthesia training (51) and urology education (53), where the value of mentor-led instruction remains, even as digital tools enhance procedural training. Our findings fully align with this perspective, advocating for the integration—not the replacement—of established pedagogical practices with AI and XR technologies. This hybrid approach ensures that digital innovation complements human interaction, reflective practice, and clinical reasoning.

While several reviews emphasize technical effectiveness, such as improved diagnostic accuracy or faster skill acquisition, systematically few address the curricular implications of these technologies. This study rectifies this deficiency by identifying the need for institutional policies, faculty training, and infrastructure development to support the sustainable integration of technology. Reviews focused on specialized training in orthopedics (52), dentistry (56), or gastroenterology (57, 58) provide valuable insights, but often limit their analysis to specific use cases within each discipline without extrapolating to broader curricular frameworks.

Our study differs methodologically from several previous works. While many are narrative or conceptual in nature, this review adopts a structured methodology using the PRISMA and PICO frameworks and focuses specifically on empirical educational interventions. By jointly examining AI-only, XR-only, and emerging AI–XR hybrid applications, it maps how these tools are currently used to support knowledge acquisition, skills performance, and learner engagement across different clinical disciplines. This synthesis underpins the curriculum-oriented recommendations developed in Sections 4.1 and 4.2, particularly in relation to competency-based design, digital literacy, and the conditions required for sustainable implementation.

This study provides an overview of the integration of AI and XR in medical education, but several limitations should be noted. First, the methodological diversity of the included studies introduces significant variability in the quality of the research and the measurement of outcomes. Many studies used cross-sectional surveys or descriptive case studies, which, while useful for exploring feasibility and user perceptions, offer limited capacity to infer causality or assess long-term educational impact. The scarcity of randomized controlled trials and longitudinal follow-ups reduces the strength of the evidence and limits the generalizability of the observed effects to other institutions and populations.

A further limitation is the variable and often small sample size; some pilot studies included fewer than ten participants. These small groups limit the statistical power of the results and restrict the conclusions that can be drawn about the scalability and reproducibility of AI and XR interventions. In addition, a substantial proportion of the included reports were explicitly described as pilot, feasibility, or development studies. These early-stage designs were deliberately retained because AI- and XR-based interventions in medical and health-professions education remain an emerging field, and many innovations are first reported in small-scale or developmental projects. However, the preliminary nature of these studies means that effect estimates should be interpreted with caution and cannot be assumed to translate directly to large-scale curricular implementation. Specifically, several studies did not provide sufficient demographic or contextual information, such as participants' educational level or prior experience with digital tools—variables relevant for evaluating learning outcomes.

A related limitation is the lack of standardized outcome measures. The tools used to assess skill acquisition, participation, or performance were heterogeneous, ranging from subjective self-assessments to simulation success rates and machine learning rankings. This inconsistency hinders comparisons between studies and compromises the possibility of conducting meta-analyses or establishing benchmark performance indicators. Additionally, many reports did not provide complete descriptive statistics or effect size estimates, which further limits the interpretability of the findings and hampers direct comparison across interventions.

Technological and institutional barriers also limited the scope of many interventions. Challenges such as insufficient digital infrastructure, limited teacher training, high implementation costs, and resistance to technological change were frequently reported. These factors not only affect the feasibility of large-scale AI/XR implementation but also influence study design, often limiting the depth or rigor of intervention evaluations. A further limitation is the heterogeneity with which the technological components themselves were reported. Some articles provide detailed descriptions of algorithms, platforms, and implementation workflows, whereas others focus mainly on curricular integration or learner outcomes and describe the underlying technology only in broad terms. This inconsistency reduces the granularity with which specific AI and XR tools can be compared across studies.

This review also primarily identified research reporting positive outcomes or improvements in learning, with fewer studies critically examining implementation failures or neutral effects. Current literature may therefore overestimate the effectiveness of these technologies in educational settings; while emerging data are promising, caution is advised when interpreting the results. More robust, comparative, and longitudinal research is needed to validate the educational value of AI and XR in diverse medical training contexts.