Artificial intelligence-assisted personal protective equipment donning and doffing training for health professions students and healthcare workers: a scoping review

Abstract

Background

Errors in donning and doffing personal protective equipment (PPE) significantly contribute to self-contamination among healthcare workers and health professions students, potentially leading to occupational exposure. Artificial intelligence (AI) offers a promising approach to enhance PPE training, but no comprehensive review of its applications and effectiveness exists.

Methods

This scoping review followed the Joanna Briggs Institute (JBI) framework and PRISMA-ScR guidelines. Four databases (PubMed, Scopus, Embase, Web of Science), grey literature, and citation searching were searched for studies published between January 2000 and November 2025. Included studies focused on AI-assisted PPE donning and doffing training for healthcare workers or health professions students.

Results

Five studies (published 2022–2025) from China (n = 2) and Australia (n = 3) met the inclusion criteria. Study designs were heterogeneous, including controlled experiment, prospective cohort, clinical cohort validation, pilot simulation study, and pre-post intervention, with sample sizes ranging from a single participant to 3382 individuals. The applied AI technologies primarily involved computer vision and machine learning, integrated into systems for real-time feedback, virtual simulation, and compliance monitoring. Evaluations suggested that AI-assisted training was associated with improved operational accuracy, with some studies reporting an increase to over 98%. One study observed a concurrent decrease in clinical infection rates, though causality cannot be established due to study design limitations.

Conclusion

AI shows strong potential to enhance PPE training through real-time feedback and personalized skill development. However, the current evidence base is limited to five studies conducted exclusively in China and Australia, which restricts the geographical generalizability of the findings. Future research should explore integrated training curricula, long-term effectiveness, and cost-efficient AI implementations.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12909-025-08498-5.

Background

In healthcare settings, personal protective equipment (PPE) acts as a critical barrier against pathogen transmission, protecting healthcare workers (HCWs) and patients [1, 2]. Its role is especially evident during outbreaks of highly infectious diseases, such as Ebola Virus Disease and Coronavirus Disease 2019 [3, 4]. However, PPE effectiveness relies on correct donning and doffing by HCWs [5, 6]. While proper donning prevents occupational exposure [7, 8], doffing errors often cause self-contamination [9–11]. Studies report self-contamination rates of 46% to 90% for various PPE types (e.g., protective clothing, gloves) across scenarios [12–14], typically due to critical errors [3, 11, 12, 15, 16], even among HCWs who perceive themselves as proficient [17].

Educational interventions enhance donning and doffing techniques, reducing self-contamination and infection risks [18, 19]. Correct PPE use is a trainable skill, despite its complexity [20, 21], and systematic training is essential [19]. Standardized proficiency lowers daily occupational exposure [7] and builds capacity for future infectious diseases [22]. For health professions students, early skill acquisition is vital. Traditional models (e.g., face-to-face demonstrations, videos, online courses) show limitations [23], with surveys indicating incomplete mastery by HCWs and students [2, 24, 25]. Face-to-face methods require significant resources, lack consistency, and are challenging during pandemics due to distancing [26–30]. Video and online options are cost-effective and safe but lack immersion and real-time feedback, leading to poor proficiency and adaptability [31]. Thus, more efficient training methods are urgently needed.

Artificial intelligence (AI) provides a promising alternative. As a computer science branch, AI builds systems simulating human intelligence [32, 33]. Recent advancements have transformed medical education paradigms [34, 35], integrating AI into scenarios like surgical simulation, nursing operations, and radiological interpretation [36–38]. AI’s features (interactive simulation, real-time feedback, personalized adjustments, and repeatability) enable its use in PPE training. Technologies like computer vision, natural language processing, and machine learning are increasingly applied to PPE scenarios, boosting efficiency and standardization. Despite these developments, a systematic literature review is lacking, leaving uncertainties about AI types, implementations, and effectiveness.

To address this gap, we conducted a scoping review. This methodology is particularly appropriate for nascent and complex research fields. Given its ability to map the breadth of a field and provide a comprehensive overview of the evidence [39], it aligns closely with our study objectives. Our review synthesizes existing AI applications in PPE training, examining contexts, outcomes, effectiveness, limitations, and challenges. It thereby generates insights to guide AI’s future in medical education and to outline pathways for optimizing PPE strategies, with the ultimate goal of enhancing healthcare workers’ protective skills.

Methods

This scoping review was conducted in accordance with the Joanna Briggs Institute (JBI) methodological framework [40] and reported following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) guidelines [41]. We provide our completed PRISMA-ScR checklist as Additional file 1. The protocol for this scoping review was registered on the Open Science Framework (https://osf.io/298er).

Research questions

This scoping review aims to systematically describe the current application status of AI in PPE donning and doffing training for health professions students and HCWs. To achieve our objectives, the following five specific research questions have been formulated.

1. What types of AI technologies are applied in PPE training?

2. How are these AI technologies applied in PPE training?

3. How is the effectiveness of AI-assisted training evaluated in existing evidence?

4. What limitations exist in the current application of AI technologies in PPE training?

5. What are the future development directions in this field?

Eligibility criteria

Explicit inclusion and exclusion criteria were developed according to the PCC framework (Population, Concept, Context) from the JBI [40] (Table 1).

Table 1.

Inclusion and exclusion criteria

	Inclusion	Exclusion
Population	Health professions students	Students from non-healthcare majors
		PPE users in non-medical fields
	Healthcare workers	Design-phase educators (non-recipients)
Concept	PPE training with defined AI mechanism	Traditional training methods
		Mere VR/AR without AI functions
		Unstated AI mechanisms
		Only involving PPE selection/management
Context	Educational institution environments	AI-assisted PPE monitoring in non-training scenarios
	Clinical practice environments	Conducted entirely in non-medical environments
	Virtual or online environments	VR/AR environment without AI functionality
Types of resources	Published: 2000−01−01 to 2025−11−16	Outside date range
	Peer-reviewed original research	Non-English language
	No geographical restrictions	Non-peer-reviewed
	No limitations on research design or quality	Non-original research
		Unavailable full text

AI artificial intelligence, VR virtual reality, AR augmented reality, PPE personal protective equipment

Population

The population included in this review comprises two groups: health professions students and HCWs. Health professions students are defined as learners enrolled in health-related programs, including but not limited to undergraduates, postgraduates, and pre-health professions students in majors such as medicine, nursing, public health, medical technology, physical therapy, occupational therapy, and physician assistant studies. HCWs cover all medical practitioners who are directly involved in patient care or may be exposed to pathogens, including registered nurses, doctors, emergency medical service personnel, public health workers, hospital technicians, and other clinical healthcare professionals.

Inclusion criteria imposed no restrictions on participants’ age, gender, years of experience, or educational background. Exclusion criteria include: students from non-healthcare majors (such as pure computer science majors), PPE users in non-medical fields (such as industrial safety personnel, law enforcement officers), and educators who only participate in training design without directly receiving AI-assisted PPE training.

Concept

The core concept focused on in this review is “AI-assisted PPE donning and doffing training,” which specifically includes the following elements:

• AI technologies: Diverse AI technologies applied in PPE training are defined as data-driven systems capable of simulating human-like decision-making. Functions such as automated guidance, real-time feedback (e.g., instant prompts for action errors), and error identification (e.g., detection of deviations in donning order through computer vision) can be realized. These technologies cover computer vision (action analysis based on convolutional neural networks), natural language processing (voice interaction with virtual tutors), machine learning (optimization of personalized training paths), and deep learning algorithms. Technologies relying solely on physical automation or simple preset rules (e.g., basic animation demonstrations) were excluded.

• Training content: Focuses on the correct procedures for donning and doffing PPE, including the usage sequence, operation skills, correction of common errors, and emergency handling of different types of PPE (e.g., gloves, masks, goggles, protective suits, etc.).

• Training modalities: Modalities include virtual reality (VR) simulations with integrated AI, augmented reality (AR)-assisted guidance, video-based motion analysis, AI-powered virtual tutor systems, and adaptive learning platforms. VR/AR systems were only considered when functioning as interactive carriers integrated with AI algorithms for intelligent decision-making (e.g., real-time feedback, error recognition).

Inclusion criteria include all studies that evaluate the application of AI technologies in PPE donning and doffing training, regardless of whether AI technologies are used as the primary training method or an auxiliary tool. Exclusion criteria include: studies that only describe traditional PPE training methods (such as face-to-face demonstrations, standard video teaching) without involving AI technologies, studies that only discuss PPE selection or management but not donning and doffing skill training, and studies that do not clearly specify the specific mechanism of action of AI technologies in training.

Context

The research contexts included in this review cover all environments and scenarios where AI-assisted PPE donning and doffing training is conducted, mainly categorized into three types:

• Educational institution environments: including classrooms, skill laboratories, or simulation centers in medical schools, nursing colleges, and other health professional education institutions.

• Clinical practice environments: encompassing actual or simulated clinical settings such as hospitals, clinics, emergency departments, and public health institutions.

• Virtual or online environments: comprising VR-based virtual campuses, remote online training platforms, and other digital learning environments.

Included studies may involve training content related to various PPE levels. Exclusion criteria include: studies on AI-assisted PPE monitoring in non-training scenarios (such as systems used only for workplace compliance checks without providing training feedback) and studies conducted entirely in non-medical environments (such as pure AI algorithm development without testing among health professional students or medical staff).

This review includes English-language, peer-reviewed original research articles published between January 1, 2000, and November 16, 2025. Conference proceedings and preprints were excluded to ensure the inclusion of studies with complete methodologies and findings that had undergone full peer review. Letters to the editor, conference abstracts, editorials, comments, reviews, and other non-original types of literature are excluded. Meanwhile, literature for which full texts cannot be obtained is excluded, with no restrictions on geographical regions or study designs.

Search strategy

The search strategy was developed in consultation with a research librarian and the research team, adhering to the JBI guidelines [42]. The following databases were searched: PubMed, Scopus, Embase, and Web of Science. Search strategies were designed and tailored for each database using a combination of keywords and subject headings related to AI-assisted PPE donning and doffing training for health professions students and HCWs.

The following search strategy was utilized to identify relevant studies: (artificial intelligence OR computer vision OR augmented reality OR Virtual Simulation OR E-Learning) AND (healthcare workers OR health professions students OR nursing students OR medical students) AND (Personal Protective Equipment) AND (training).

The complete search strategy for each database is provided in Additional file 2. The initial database search was conducted on July 10, 2025. Grey literature searches through Google Scholar, OpenGrey, and ProQuest were carried out on July 15, 2025. An updated search was performed on November 16, 2025, to include any newly published literature.

Study selection

All identified articles were imported into the reference management software EndNote (Version X9), with duplicate documents automatically removed. Two reviewers (LV and XQ) independently screened the titles and abstracts of the identified articles, and conducted full-text assessments for those that potentially met the eligibility criteria. The reasons for inclusion and exclusion were documented for each article. The consistency of screening results between the two reviewers was evaluated using Cohen’s kappa (κ = 0.85). Any disagreements during the title/abstract or full-text screening stage were resolved through discussion between the two reviewers. If consensus could not be reached, a third reviewer (WPP) was consulted to make the final decision. Furthermore, no quality assessment of the included studies was performed, as this is not a mandatory component of scoping reviews [43]. An overview of the study selection process is presented in the PRISMA flow diagram (Fig. 1).

Fig. 1.

PRISMA Flow Diagram of Study Selection. Flowchart illustrating the identification, screening, and selection process of studies included in this scoping review. A total of 281 records were identified through database searches, with 58 duplicates removed prior to screening. After screening 223 titles and abstracts, 28 full texts were assessed for eligibility, resulting in 3 included studies. An additional 95 records were identified through other methods (72 from grey literature and 23 through citation searching), with 28 duplicates removed. After screening 67 titles and abstracts and assessing 17 full texts, 2 additional studies were included. In total, five studies met the eligibility criteria and were included in the final review. Abbreviations: AI = Artificial Intelligence, PRISMA = Preferred Reporting Items for Systematic Reviews and Meta-Analyses

Data extraction

A standardized data extraction form was developed by the research team, and all data were recorded in Microsoft Excel. The two independent reviewers mentioned above (LV and XQ) continued to extract data from the included full texts, with discrepancies resolved through team discussions. The extracted data were randomly checked by another two reviewers (WPP and CMH) to ensure accuracy and consistency. All authors reviewed and reached a consensus on the final articles for data extraction. The data extracted included the following information: (1) Bibliographic details (authors, publication year, country); (2) Study characteristics (study design, aim, sample size, participant demographics); (3) Intervention details (AI technology used, training content, delivery mode, application scenario); (4) Evaluation methods (outcomes measured, assessment tools, follow-up duration); and (5) Key findings (quantitative results on accuracy, efficiency, user feedback).

Data analysis

In accordance with the recommendations proposed by Westphn et al. (2021) [44], a mixed-methods approach combining descriptive statistics and thematic analysis was employed to analyze the extracted data. The analytical process was carried out collaboratively by the research team until consensus was achieved. Firstly, descriptive statistics were conducted. Basic characteristics of the included studies were calculated, including year of publication, geographical region, type of study design, sample size range, and composition of participant types. These statistics were presented using frequencies and percentages. Secondly, thematic classification analysis was carried out following the framework proposed by Braun & Clarke [45]. Focusing on the five research questions, analyses were performed as follows: categorizing the types of AI technologies; summarizing application scenarios and operational workflows of AI technologies; synthesizing and classifying training outcome evaluation metrics to illustrate overall trends; quantifying the frequency of limitations across dimensions such as technical maturity; and identifying future research directions. Finally, the analysis results were integrated and presented in the form of tables and charts.

Results

Screening results

A total of 376 records were identified across all sources. From database searches, 281 records were identified. After the removal of 58 duplicates, 223 records were retained for title and abstract screening, during which 195 records were excluded. This left 28 records for full-text retrieval, all of which were successfully retrieved. Following eligibility assessment, 25 articles were excluded, resulting in 3 studies being retained from databases. Additionally, 95 records were identified from other methods. After the removal of 28 duplicates, 67 records were screened by title and abstract, leading to 45 exclusions. Of the remaining 22 records sought for retrieval, 17 were retrieved. After eligibility assessment, 15 articles were excluded, resulting in 2 studies being retained from other methods. Finally, five studies were included in this review.

Study characteristics

The characteristics of the five included studies are summarized in Table 2. Published between 2022 and 2025, these studies originated from China (n = 2) and Australia (n = 3), indicating the nascent and geographically concentrated nature of this field. Study designs were heterogeneous, including one controlled experiment [46], one prospective cohort study [47], one clinical cohort validation study [48], and two pilot studies (one simulation-based [49] and one hospital-based intervention study [50]). Sample sizes varied substantially, ranging from a single operator in a feasibility simulation [49] to a large cohort of 3382 healthcare workers [50], reflecting a spectrum from technical validation to larger-scale implementation. Participants included both practicing HCWs and health professions students, confirming the relevance of AI-assisted training across the professional continuum. All studies aimed to evaluate the effectiveness, real-time guidance capability, and potential impact on infection control of their respective AI-assisted training systems.

Table 2.

Characteristics of included articles

References Number	Authors (Year)	Country; Setting	Study Design	Population (Sample Size)	AI Technology/Intervention	Comparator	Key Outcomes (with denominators & time frame)	Follow-up Duration
[46]	Qin et al. (2022)	China; University	Controlled Experiment (4 groups)	Inexperienced medical students (n = 48); experienced HCWs (n = 48)	Gesture-Control System (MediaPipe)	Live demo, Virtual simulation, No training	Inexperienced medical students (Immediate post-training): 1. Donning Accuracy: Gesture system − 93.53%, Live demo − 85.21%, Virtual simulation − 83.40%, No training − 69.44%. 2. Doffing Accuracy: Gesture system − 80.00%, Live demo − 75.00%, Virtual simulation − 75.83%, No training − 47.19%. Experienced HCWs (Immediate post-training): 1. Donning Accuracy: Gesture system − 99.08%, Live demo − 94.50%, Virtual simulation − 95.42%, No training − 95.39%. 2. Doffing Accuracy: Gesture system − 96.67%, Live demo − 90.83%, Virtual simulation − 90.00%, No training − 80.83%.	1 test session (post-training)
[47]	Preda et al. (2025)	Australia; Hospital	Prospective Cohort (Pre-Post)	HCWs (n = 293)	SXR AI-PPE Platform (Guided vs. Unguided Modes)	Baseline (Guided) Performance	Longitudinal Subgroup (n = 20): 1. Pass Rate after 2 sessions: 100% 2. Self-efficacy (Confident): 33% (Pre) → 80% (Post) 3. Time Reduction: Donning − 7.2%; Doffing − 11.3%	6 months (3-month intervals)
[48]	Preda et al. (2022)	Australia; Hospital	Clinical Cohort Validation	HCWs (n = 74)	SXR AI-PPE Platform (Optical Classifiers)	Double-Buddy Gold Standard	Sensitivity (Immediate assessment): 1. Donning: 85.3% (p < 0.01) 2. Doffing: 98.9% (p = 0.125) 3. Buddy Correction Rate: 3.8% ± 1.5%	Single session
[49]	Segal et al. (2023)	Australia; Simulation	Pilot Simulation Study	1 Operator (30 Scenarios)	Blue Mirror Software (Computer Vision) + Remote Buddy	On-site Buddy	Accuracy: 1. Human-AI System: 100% (195/195 steps) 2. On-site Buddy: 99% 3. AI Autonomy: 89% (173/195 steps) 4. Inter-rater agreement: κ = 0.97	Single session
[50]	Huang et al. (2023)	China; Hospital	Pre-Post Intervention	HCWs (n = 3382)	AITMS System (Kinect-based Motion Capture)	Pre-AITMS Period	Operational Accuracy: 1. Pre-AITMS: 52.15% (85/163) 2. Post-AITMS: 98.14% (3159/3219) 3. Statistical Test: χ²=834.35, p < 0.001	2-year intervention period

HCWs healthcare workers, SXR Surgical XR, AI artificial intelligence, PPE personal protective equipment, AITMS AI-based training and monitoring system

Types and core functions of AI technologies

The AI technologies employed in the included studies were primarily centered on computer vision and machine learning, integrated with real-time feedback mechanisms to support PPE training. A summary of the five key AI systems, their core technologies, and primary functions is provided in Fig. 2. For example, a gesture-controlled system was developed based on the MediaPipe Toolkit [51], utilizing hand detection and gesture recognition algorithms. Video guidance was controlled via “OK” gestures (e.g., right-hand “OK” to proceed to the next step, left-hand “OK” to return to the previous step) [46]. The Surgical XR (SXR) AI-PPE Platform employed an optical classifier to identify PPE items (e.g., protective gowns, masks, goggles) and provided step-by-step audiovisual prompts through a guided mode [47]. Both the AI-PPE Platform and the Blue Mirror Software combined computer vision with a human-machine collaboration mode. The former utilized a dual-partner system to verify accuracy, while the latter was designed such that 89% of the steps were autonomously monitored by AI, with the remaining steps corrected by a remote human partner [48, 49]. The AITMS System used Kinect-based technology to capture human body posture. This was combined with a machine learning model to evaluate standardized procedures, enabling real-time voice-activated error correction [50].

Fig. 2.

Types and Core Functions of AI Technologies in PPE Training Schematic overview of five key AI systems (SXR AI-PPE Platform, Gesture-Controlled System, AI-PPE Platform, Blue Mirror Software, AITMS System) used in PPE training, including their core AI technologies and primary functions. Abbreviations: AI = Artificial Intelligence, PPE = Personal Protective Equipment

The application scenarios of AI in PPE training

The application scenarios of AI systems were categorized into three primary types: training, real-time guidance, and routine monitoring. Some systems were designed with dual-mode functionality to accommodate different needs. In training mode, the SXR AI-PPE Platform and the AITMS System were used to provide novices with theoretical knowledge and standardized video demonstrations. Learning outcomes were assessed through unguided evaluations [47, 50]. In real-time guidance mode, the Gesture-Control System and the AI-PPE Platform supported HCWs by providing gesture-based or AI-prompted corrections during actual donning and doffing procedures (e.g., adjusting the nose clip of a mask, verifying the order of fastening protective gown ties) [46, 48]. In routine monitoring mode, the Blue Mirror Software and the AITMS System were deployed for compliance checks of PPE use in clinical settings. Alerts were triggered for errors such as hair exposure or contact with contaminated zones [49, 50]. Some systems (e.g., SXR AI-PPE, AITMS) offered mode-switching capability via mobile devices, enabling both daily training and clinical operation monitoring [47, 50].

Evaluation of AI-Assisted training effectiveness

The findings demonstrated that AI-assisted training was associated with significant improvements in the accuracy, efficiency, and self-efficacy of PPE donning and doffing procedures. In terms of accuracy, the Gesture-Control System enabled inexperienced users to achieve a donning accuracy rate of 93.53% and a doffing accuracy rate of 80.00%, which were significantly higher than those of traditional virtual simulation and on-site demonstration [46]. Following two guided training sessions with the SXR AI-PPE Platform, a participant accuracy rate of 100% was achieved and subsequently maintained [47]. The AITMS system was associated with an increase in operational accuracy (52.15% to 98.14%), and one study reported a concurrent decrease in hospital infection rates (1.39% to 0.38%); however, the observational nature of the study precludes causal inference [50].

Discussion

This is the first scoping review to systematically integrate evidence on the application of AI in training health professional students and HCWs in PPE donning and doffing. Analysis of the five included studies indicates that technologies primarily using computer vision (action recognition) and machine learning (personalized path optimization) were reported to improve PPE training effectiveness through real-time feedback, virtual simulation, and compliance monitoring. Given the small number of included studies and their methodological heterogeneity, all findings should be interpreted as associations rather than causal effects, and the overall evidence remains preliminary.

Cognitive load theory: AI Real-Time feedback optimizes working memory burden

Cognitive Load Theory posits that learning efficiency declines when working memory is overloaded [52]. Traditional PPE training often induces such overload by requiring learners to simultaneously memorize multi-step procedures and operational details, consuming cognitive resources with extraneous information [31].

AI-assisted training has been proposed to address this issue by managing cognitive load. It reduces extraneous load through real-time, automated feedback (e.g., voice prompts for error correction), freeing working memory to focus on action execution rather than recall [50]. Concurrently, it manages intrinsic load by decomposing complex procedures into discrete, validated micro-steps, preventing informational overwhelm [53, 54]. This approach is supported by physiological evidence; studies using electroencephalogram have shown significantly lower cognitive load in the frontal lobe among AI-guided trainees compared to those in conventional training groups (p < 0.01) [55]. By externalizing procedural knowledge, AI effectively offloads working memory, enabling learners to concentrate on critical decision-making.

Dreyfus model: AI simulation training accelerates the Novice-to-Expert transition

The Dreyfus model describes skill acquisition as a progression from rule-based novice to intuitive expert [56]. Findings suggest that AI-assisted training may support learners during this transition through targeted interventions at key stages.

For the advanced beginner to competent transition, AI generates diverse clinical scenarios (e.g., low-light donning, sudden contamination). This helps learners move beyond fixed rules by recognizing contextual cues, as demonstrated by the SXR AI-PPE platform, where varied training was associated with sustained 100% accuracy [47].

For the competent to proficient transition, AI provides personalized feedback based on systematic error analysis. This promotes deeper “goal-operation” associations rather than mechanical rule-following, aligning with the development of goal-oriented thinking at this stage.

Finally, for the proficient to expert transition, AI supports unlimited repetitive practice. This fosters the automation of procedural steps, reducing conscious effort. For instance, one human-machine collaboration system achieved a zero error rate through extensive autonomous training sessions, underscoring the role of repetition in developing expert-level automaticity [49].

Comprehensive comparison of AI, VR, and E-Learning in PPE training

AI, VR, and e-learning are prominent technologies in digital medical education, all of which have been applied in PPE training. Based on the limited evidence available, including the studies in this review and related work on VR and e-learning, we synthesize and compare their technical characteristics, efficacy, and applicability. It is important to note that this comparison is preliminary due to the small and heterogeneous evidence base, particularly for AI. The following analysis aims to delineate their distinct, and often complementary, strengths.

Comparison of technical characteristics

AI-assisted training is driven by computer vision (e.g., OpenPose) and machine learning. Through motion sensors (Kinect/RFID), millimeter-level operational errors (2 mm) are captured in real time, and voice or visual error-correction instructions (e.g., “Incorrect glove doffing sequence”) are dynamically generated [47]. The technical core of such systems lies in real-time perception and dynamic feedback. No preset content is required, and the interactive logic can be dynamically adjusted in response to the trainee’s performance.

VR relies on head-mounted displays to construct high-fidelity virtual environments, through which details such as the deformation of N95 respirators are realistically replicated using physics engines (e.g., GGX algorithm combined with XPBD cloth simulation). However, due to hardware limitations, the tracking accuracy of complex hand gestures is often inadequate. Lower-end devices exhibit frame rates of only 22–29 fps, and approximately 15% of users suffer from dizziness [57, 58]. Its technical core is the construction of immersive scenes, with the sense of immersion enhanced through 360 scenario simulation.

• Learning is typically delivered through online platforms, such as those built on content management systems like Joomla. Instructional interactivity is achieved through the incorporation of gamified designs, such as drag-and-drop sequencing of PPE components and contextual multiple-choice tasks. For example, Gamified e-learning modules were developed using Articulate Storyline 3 and were designed to be accessible across smartphones and computers. The approach is centered around preset content and standardized interactions [59].

Differences in training efficacy

Knowledge transfer and basic skills

No significant advantage was observed for AI-assisted training in knowledge transfer, but its real-time error correction capability is more conducive to the conversion of knowledge into motor memory. In contrast, VR demonstrates a unique capacity to enhance memory encoding through immersive scenarios, proving particularly suitable for high-risk situational awareness training. However, its current limitations include the inability to fully substitute tactile feedback, a factor identified by 59.3% of practitioners as a constraint on operational accuracy [60]. E-learning has demonstrated effectiveness in the standardized delivery of foundational PPE knowledge, as evidenced by significant improvements in correct PPE selection rates [59, 61, 62]. However, its utility appears more limited for acquiring the more complex practical skills of donning and doffing sequences, where studies have shown it to be less effective when used in isolation [61].

Practical skills and complex tasks

Within the limited studies reviewed, AI-assisted training suggested potential for targeted reinforcement of weak procedural areas through error pattern analysis. For example, one study reported a correct performance rate of 89% for containment-based doffing of protective clothing, and another observed a 52% reduction in decision-making time for high-level operational skills [47, 49]. These initial findings point to AI’s promise in optimizing complex psychomotor skills, though they require confirmation in larger trials. In VR-based training, correct performance rates ranging from 90% to 100% were attained in basic donning and doffing sequences. However, error rates reached up to 37% in advanced skills such as contamination prevention during containment-based doffing [28, 60]. E-learning demonstrated limited effectiveness in teaching practical operational skills such as donning and doffing sequences. Only 7.7% of trainees were able to correctly perform all steps, and none could accurately describe the doffing sequence [61]. It is therefore recommended that e-learning be combined with face-to-face teaching, such as the Peyton four-step method, to improve long-term retention [63].

Cost and application scenarios

The core cost of e-learning is concentrated in content development, including specialized tools and expert instructional design teams. Once developed, the content can be reused, but updates depend on expert teams. This limitation was highlighted by 74% of learners who requested greater diversity in training modules [59, 61]. VR requires high hardware investment (e.g., Oculus Quest 2 costs $400 per unit) and incurs high content update costs [64]. In contrast, AI sensor network systems can be deployed at less than one-fifth the cost of VR setups. Furthermore, AI platforms support simultaneous multi-learner monitoring, resulting in lower long-term operational costs [65].

• E-learning is best suited for standardized introductory training, such as teaching fundamental PPE selection principles. Support is provided for cross-platform collaboration, enabling up to 20 users to access training modules simultaneously [59, 62]. VR is particularly effective for high-risk situational simulations, including contamination emergency response. Immersive training is delivered through platforms such as the MAGES engine, which integrates instructor demonstration with trainee practice [65]. AI is optimally employed for personalized practical guidance and compliance auditing. Remote error correction and large-scale performance evaluation are enabled through real-time data-sharing mechanisms.

• In summary, the preliminary evidence, while limited, allows for the identification of distinct and complementary profiles for AI, VR, and E-learning within PPE training. E-learning is recognized for its effectiveness in foundational knowledge transfer and standardized content dissemination. VR demonstrates particular strength in high-risk situational immersion and team collaboration exercises. Meanwhile, AI, based on early-stage studies, shows potential advantages in real-time performance optimization and personalized skill reinforcement. Acknowledging that this landscape is evolving, these technologies should not be viewed as mutually exclusive but as components of a potential blended learning ecosystem.

Strengths and limitations of the study

To the best of our knowledge, this represents the first scoping review focused on the application of AI in PPE training. The scoping review was rigorously conducted following the JBI framework and the PRISMA-ScR guidelines, ensuring methodological transparency and reliability. In addition, independent screening and data extraction by two authors reduce the risk of potential bias.

Despite these strengths, several limitations must be acknowledged. First, and most notably, the evidence base for this review is constrained by the nascent state of the field, resulting in the inclusion of only a small number of studies (n = 5). A critical limitation stemming from this is the pronounced geographical concentration of the included research, with all studies originating from only two countries (China and Australia). This lack of geographical diversity significantly limits the generalizability of our findings. The applicability of these AI technologies, their effectiveness, and their implementation challenges may vary considerably across different healthcare systems, cultural contexts, and resource settings. Therefore, our results and discussions should be interpreted primarily within the context of the represented regions.

Second, non-English publications were excluded, which may have restricted the scope of available evidence. Third, a noticeable technological bias was observed: current research predominantly emphasizes computer vision (e.g., Kinect-based motion capture) and basic machine learning algorithms, with insufficient exploration of more complex AI technologies such as natural language processing for real-time trainee support. Fourth, comparative studies are relatively lacking. Although differences among AI, VR, and e-learning were analyzed, there is a lack of head-to-head comparisons under consistent conditions (e.g., same population and evaluation metrics). As a result, some conclusions remain dependent on indirect inference.

Finally, in line with the purpose of a scoping review to map the available evidence regardless of methodological quality, we did not perform a formal quality appraisal of the included studies. However, this decision necessitates a cautious interpretation of the findings. The methodological rigor of the five included studies varied. For instance, the evidence base consists predominantly of controlled experiment, pilot simulation, and pre-post intervention [46, 49, 50], which are essential for proof-of-concept but often feature small sample sizes or lack controlled comparisons. Furthermore, the reliance on self-reported outcomes or unblinded assessments in some studies may introduce performance and detection bias [47]. Therefore, while the initial results regarding the feasibility and potential effectiveness of AI-assisted PPE training are highly promising, the current evidence must be considered preliminary. The positive findings require confirmation through more robust study designs, such as larger-scale randomized controlled trials with long-term follow-up and objective outcome measures.

Practical implications

The included studies suggest the potential value of AI-assisted PPE training across three key aspects. Firstly, it shows potential in enhancing operational accuracy. Millimeter-level errors (2 mm) are captured in real time through computer vision and sensors, while targeted corrective feedback is simultaneously generated. This approach appears to effectively address the shortcomings of traditional training in fine motor skill standardization. For example, correct performance rates for high-level skills were increased to 89%, significantly reducing the error rates of 27.3%–34.1% observed in conventional training [31, 47, 49].

Secondly, AI enables personalized skill progression. For novices, cognitive load is reduced through micro-task decomposition, resulting in a 19% reduction in frontal lobe load (p < 0.01) [55]. For advanced learners, high-frequency error steps are reinforced through targeted training, facilitating the transition from rule-based execution to adaptive response. For experts, varied scenarios are generated to maintain skill acuity. These strategies align closely with the developmental stages outlined in the Dreyfus model.

Thirdly, AI is suitable for large-scale implementation. The cost of sensor networks is only one-fifth that of VR systems [65]. Multi-trainee synchronous monitoring is supported, and all performance data are traceable. This enables both large-scale training deployment and compliance auditing, making it particularly applicable for standardized implementation in resource-limited settings.

While the improvement in operational accuracy is a key metric for training efficacy, its ultimate value is determined by the translation into improved clinical outcomes, particularly the reduction of healthcare-associated infections. Among the included studies, only Huang et al. (2023) [50] directly investigated this relationship, reporting an association between the implementation of their AI system and a concurrent decrease in nosocomial infection rates (from 1.39% to 0.38%). However, as an observational study, this finding must be interpreted with caution. The temporal correlation does not establish causality, and confounding factors (e.g., concurrent changes in infection control protocols or community prevalence of infections) could not be controlled [66].

The remaining four studies [46–49] did not assess the impact of AI-assisted training on clinical infection rates. Their primary endpoints were confined to proximal outcomes, such as procedural accuracy, compliance, time efficiency, and self-efficacy. This focus on skill acquisition and short-term performance is consistent with the early-stage, proof-of-concept nature of most current AI applications in PPE training. The gap between demonstrating proficiency in a controlled or simulated environment and effecting a measurable change in population-level infection rates is well-documented in the broader literature on clinical training [67]. Bridging this gap requires study designs that not only validate the training tool but also directly link its use to long-term, clinically significant endpoints under real-world conditions.

Directions for future research

Based on the preliminary findings and limitations of the current evidence base, six key directions are proposed for further investigation.

First, the development of integrated training curricula should be explored through randomized controlled trials combining e-learning, VR and AI. The optimal connection mode of the three stages, which form a sequential pipeline from e-learning (knowledge transfer) to VR (immersive simulation) and finally to AI (hands-on skill adaptation) as illustrated in Fig. 3, should be clarified. Subsequently, simulated clinical assessments should be used to evaluate their synergistic effects on knowledge retention and skill transfer rates.

Fig. 3.

Integrated PPE Training Curricula Combining E-Learning, VR, and AI Diagram of the three-phase hybrid training model: Phase 1 (E-Learning) for foundational knowledge transfer, Phase 2 (VR) for immersive scenario simulation, and Phase 3 (AI) for hands-on skill adaptation. Abbreviations: VR = Virtual Reality, AI = Artificial Intelligence

Second, long-term effectiveness and real-world validation should be expanded. Studies tracking the impact of AI-assisted training on PPE-related infection rates in clinical practice are recommended. The pathway from training efficacy to clinical outcomes should be further verified, particularly through the implementation of low-cost AI solutions (e.g., smartphone camera-based motion recognition) in resource-limited settings such as primary care institutions.

Third, technological innovation should be promoted through the development of multimodal AI systems that integrate motion recognition and physiological signal monitoring. This enables the synchronous assessment of both operational errors and psychological stress.

Fourth, a standardized evaluation framework must be established. Effectiveness indicators for AI training, such as error correction accuracy and cognitive load reduction, must be clearly defined. This will provide a basis for horizontal comparisons between different studies.

Fifth, bridging research and practice is essential for realizing the equitable and immediate impact of AI-assisted training. Future work must prioritize the development and validation of low-cost, offline-capable AI solutions (e.g., leveraging smartphone cameras) for primary care institutions in low- and middle-income countries. This includes investigating strategies to overcome infrastructural barriers, such as intermittent internet connectivity, and ensuring the cultural appropriateness of training content. Concurrently, for immediate integration, educators should be encouraged to adopt these technologies as supplementary tools within existing PPE curricula. This can be achieved by deploying AI for self-directed practice in skills labs, utilizing its objective data for formative assessment, and providing targeted remediation for struggling learners, thereby building a foundation for the seamless adoption of more advanced solutions in the future.

Finally, expanding the scope and diversity of evidence is critical to address the current limitations of geographical concentration and small sample sizes. Future systematic reviews and primary studies should actively incorporate regional databases, grey literature, and publications in languages other than English. Including varied study designs, such as qualitative investigations and detailed case reports, will enrich the understanding of implementation contexts and user experiences. Furthermore, promoting multicenter studies and international collaborations is strongly encouraged to enhance geographic representation and validate the findings across different healthcare systems and educational cultures.

Conclusion

Based on the limited but promising evidence from five included studies, this scoping review suggests a promising potential for AI-assisted training in PPE donning and doffing. The limited available studies associate mechanisms like real-time feedback and personalized reinforcement with improvements in practical skill accuracy and reductions in cognitive load. Our analysis found that current applications are predominantly driven by computer vision and machine learning technologies, integrated into interactive training, real-time guidance, and compliance monitoring systems. Initial evidence indicates that these AI-assisted approaches may enhance operational accuracy, efficiency, and self-efficacy compared to traditional methods.

However, the field remains nascent. The findings are derived from a small number of heterogeneous studies, and other limitations include a narrow technological focus, geographical concentration, and a lack of robust comparative trials. Compared with e-learning and VR, AI may be more suitable as an advanced training tool, with the three forming a complementary relationship. The hybrid model of “e-learning + VR + AI” is proposed as a potentially efficient and scalable solution for future PPE training. With continued technological iteration and validation through larger-scale studies, AI-assisted training could evolve into a valuable component in fortifying the protective capabilities of health professions students and HCWs. Such an evolution could contribute to reducing occupational exposure risks and strengthening preparedness for future public health emergencies.

Abbreviations

AI

Artificial Intelligence

AR

Augmented Reality

HCWs

Healthcare Workers

JBI

Joanna Briggs Institute

PCC

Population, Concept, Context

PPE

Personal Protective Equipment

PRISMA-ScR

Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews

SXR

Surgical XR

VR

Virtual Reality

Biographies

Yu Lv

is a senior lecturer and academic developer at the School of Medicine, University of Electronic Science and Technology of China, Chengdu, China.

Minhong Cai

is a Professor of Medicine at the School of Medicine, University of Electronic Science and Technology of China and a specialist in hospital infection prevention and control at Sichuan Provincial People’s Hospital, Chengdu, China.

Qian Xiang

is a senior lecturer in medical education and researcher at the School of Medicine, University of Electronic Science and Technology of China, Chengdu, China. He is the Director of Hospital Infection Prevention and Control Center and Head of Hospital Infection Course.

Pingping Wang

is an assistant lecturer in medical education at the School of Medicine, University of Electronic Science and Technology of China and an attending physician at Sichuan Provincial People’s Hospital, Chengdu, China.

Authors’ contributions

WPP conceived the study, developed the review protocol, and was responsible for the literature search. LV and XQ performed study screening and data extraction. WPP and CMH supervised all stages of the study. WPP and LV drafted the manuscript. All authors contributed to revising and improving the draft. All authors have read and approved the final manuscript.

Funding

This study was financially supported by the National Natural Science Foundation of China (No. 21976046 & No. 22476032), Science and Technology Department of Sichuan Province (investment ID: 2023NSFSC0534).

Data availability

The data that supports the results and findings of this scoping review can be found in the manuscript. Any other data from the current study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.