2.2.1 Population.

The population in this study is defined as HCWs (e.g., physicians, nurses, medical technicians) performing hand hygiene behaviors in healthcare institutions, as well as healthy volunteers or other individuals simulating the roles of HCWs.

2.2.2 Concept.

The core concept of this review centers on AI technologies applied to the automated monitoring, recognition, or evaluation of the aforementioned hand hygiene behaviors. This encompasses, but is not limited to, systems utilizing computer vision, deep learning, sensor fusion, or related technologies.

2.2.3 Context.

The review includes studies conducted in real or simulated healthcare environments, such as hospital wards (e.g., ICUs, general wards), operating rooms, emergency departments, or other acute and chronic care settings.

2.2.4 Types of source.

This review includes English-language peer-reviewed original research articles published between January 1, 2000, and September 30, 2025. Letters to the editor, conference abstracts, editorials, commentaries, review articles, and other non-original publications are excluded. Studies for which full text cannot be accessed will also be excluded. No restrictions are applied regarding geographical region or study design.

3.2.1 Publication year and geographical distribution.

The included studies span publication years from 2007 to 2025, with 93.3% (42/45) published after 2015, reflecting a marked increase in research activity in this field over the past decade. Geographically, the majority of studies originated from the United States [30,34,49,59–61,63,65,70,77,80,81] (n = 12, 26.7%), followed by China [42,47,48,51,52,57,76] (n = 7, 15.6%). Australia [28,75,79], India [56,72,74], and Spain [58,66,67] each contributed three studies. Italy [27,62], Ireland [45,78], Turkey [50,73], Germany [53,64], and Latvia [54,68] each provided two studies. Single studies were contributed by South Korea [43], Saudi Arabia [44], the United Kingdom [35], Colombia [46], Japan [55], Vietnam [69], and Singapore [71]. These findings indicate a global distribution of research in this field, though contributions are exclusively from high- and middle-income countries, with no representation from low-income countries. Detailed distributions of publication years and geographical origins are presented in Fig 2.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Publication year and geographical distribution.

Figure (a) presents the distribution of included studies by year of publication. Figure (b) illustrates the geographical distribution of the study locations, with the bar heights representing the number of publications originating from each country. Abbreviations: USA, United States of America; UK, United Kingdom.

https://doi.org/10.1371/journal.pone.0347683.g002

3.2.2 Study design and setting distribution.

Among the included studies, the distribution of study designs was as follows: experimental studies accounted for 37 studies (82.2%), quasi-experimental studies for 5 (11.1%), technical validation studies for 2 (4.4%), and diagnostic accuracy studies for 1 (2.2%). The healthcare settings covered included general wards (n = 35, 77.8%), ICUs (n = 13, 28.9%), emergency departments (n = 3, 6.7%), operating rooms (n = 2, 4.4%), dental clinics (n = 1, 2.2%), and general clinics (n = 1, 2.2%). Several studies encompassed multiple clinical environments, resulting in percentage totals exceeding 100%. High-risk infection settings, including ICUs, emergency departments, and operating rooms, were focal points of research attention. The distribution of study designs and settings is detailed in Table 2.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Study design/validation categories and settings of included studies.

https://doi.org/10.1371/journal.pone.0347683.t002

3.2.3 Study participants and sample characteristics.

Among the 45 included studies, at least 378 clinical HCWs, including physicians, nurses, and medical students, were explicitly reported as participants [42,45,47,65,68,73,76,77]. Additionally, some studies utilized healthy volunteers to simulate clinical HCWs [27,28,50,57–59,61,66,75], whereas other studies did not report the exact number of participants, reporting only overall sample sizes [30,34,35,43,44,46,48,49,51–56,60,62–64,67,69–72,74,78–81]. Reported participant numbers and sample characteristics varied substantially across studies. Video and image datasets also demonstrated considerable variation, spanning thousands to tens of thousands of frames across different studies [42,47,58,68,77]. Small-scale experimental investigations primarily focused on algorithm validation. Examples included 8 nurses generating 12,036 samples in an ICU setting [76], 22 HCWs recruited for cumulative observation in the Geneactiv Shadowing project [77], and 6 participants providing continuous 7-day data documenting 365 hand hygiene events via low-cost wearable sensors [58]. Moderate-scale clinical evaluations included a multi-camera system deployed across five departments capturing 165 video segments from 55 physicians and nurses [47]. Large-scale clinical validation was demonstrated in a burn center study monitoring 41 HCWs during 20,095 hand hygiene opportunities, documenting 15,202 compliance events [73]. The ICU-MH database contained 74,471 image frames from 20 HCWs [42]. In non-clinical technical development, one investigation utilized 32,471 publicly annotated video sequences for algorithm training [72], while another study conducted comparative experiments with 72 medical personnel at a medical school [68].

3.2.4 Study maturity and validation context.

To improve interpretability, the included studies were also considered in terms of study maturity and validation context. Overall, the evidence base was dominated by early-stage investigations, with most studies focusing on technical development, controlled validation, or limited pilot testing rather than sustained real-world implementation. As shown in Table 2 and Table 3, experimental studies accounted for the majority of included articles, whereas only a small proportion were quasi-experimental, diagnostic accuracy, or technical validation studies. In addition, the validation context presented in Table 3 indicates that most studies were still at the stage of technical development, controlled validation, or limited pilot clinical evaluation. Only a few studies reflected broader real-world implementation. Sample sizes and evaluation conditions also varied substantially: several studies involved small numbers of participants or highly controlled datasets for algorithm development, while only a few studies reported larger-scale clinical evaluations or deployment in routine care settings. Accordingly, the reported performance of AI-based hand hygiene monitoring systems should be interpreted in light of study maturity, sample size, and validation context, and not assumed to reflect equivalent levels of clinical readiness or generalizability across studies.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Technical pathways, characteristics, and validation context / study maturity of AI-based hand hygiene compliance monitoring.

https://doi.org/10.1371/journal.pone.0347683.t003

In addition, important heterogeneity was observed in how study outcomes, reference standards, and validation strategies were defined across the included literature. Some studies evaluated hand hygiene event detection or opportunity recognition, whereas others focused on procedural step recognition, duration compliance, technique quality, or post-implementation compliance improvement. Reference standards also varied, including direct human observation, expert video annotation, sensor-derived event logs, and internally constructed technical benchmarks. Validation strategies ranged from laboratory-based algorithm development and internal dataset testing to pilot clinical evaluations and limited real-world implementation. This heterogeneity further limits direct comparison of reported performance metrics across studies and reinforces the need to interpret accuracy estimates in light of study maturity, evaluation target, and validation context.

To facilitate interpretation of the reported performance measures, Table 4 summarizes the included studies according to validation context and evidence maturity, distinguishing technical development/laboratory validation, pilot clinical evaluation, and broader real-world implementation. This framework provides additional context for understanding why reported accuracy and related metrics should not be directly compared across studies or interpreted as equivalent indicators of clinical readiness.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Framework for interpreting reported performance metrics according to validation context and evidence maturity.

https://doi.org/10.1371/journal.pone.0347683.t004

3.3.1 Computer vision-based systems (24 studies, 53.3%).

As the predominant technical approach, these systems utilize cameras for non-contact monitoring, offering the key advantage of objectively recording hand hygiene compliance, movement accuracy, and duration. Functionally, such systems comprehensively cover hand hygiene event detection (24 studies), technique classification (17 studies), duration estimation (15 studies), and real-time feedback (10 studies), thereby providing multi-dimensional data for healthcare-associated infection surveillance.

System performance is closely linked to potential clinical applicability, particularly in studies approaching clinical application: personalized models trained on specific HCWs can achieve high recognition accuracy, such as 95% in ICU settings [76], although this level of performance was reported under setting-specific conditions and does not imply comparable generalizability across users or institutions. In contrast, generalizable models designed for cross-user application exhibit reduced performance due to variations in hand size and washing habits, with accuracy potentially declining to 56% [76]. This finding underscores the importance of model adaptability for future real-world deployment rather than indicating equivalent readiness across all study contexts.

Moreover, privacy concerns represent a significant challenge in clinical implementation. Multi-camera fusion [47] may help reduce occlusion in complex clinical environments, with reported recognition gains of 12%–18%.

3.3.2 Wearable sensor-based systems (11 studies, 24.4%).

These systems are suitable for mobile clinical settings with limited camera coverage, such as ICUs and general wards. Their advantages include portability and extended battery life, with low-power designs enabling continuous operation for 7–14 days [58]. The primary functions of these systems focus on estimating hand hygiene duration (11 studies) and detecting the occurrence of the action (8 studies).

A key limitation for clinical application is their relatively low specificity, which renders them susceptible to false alarms triggered by similar hand movements, such as writing or operating medical instruments. This results in a specificity that is 5%–10% lower than that of computer vision-based systems [77]. To improve monitoring accuracy, studies have demonstrated that dual-wrist configurations [77] capture bimanual coordination more effectively than single-wrist setups, increasing recognition accuracy by 8%–15%. Furthermore, some systems, when integrated with GPS positioning, have achieved linked monitoring of “hand hygiene location and behavior,” reporting false alarm rates below 5% [80]. This capability provides valuable managerial insights for analyzing the relationship between compliance rates and specific clinical locations.

3.3.3 Radar/radio frequency-based systems (4 studies, 8.9%).

This category of technology provides a privacy-centric alternative for non-contact monitoring, as its operational principle does not involve capturing visual data. This attribute makes it particularly suitable for deployment in privacy-sensitive environments such as patient rooms. For instance, one implementation employs gas sensors [53] to support the verification of hand hygiene events by rapidly measuring the volatile alcohol concentration emitted from hand rub solutions, achieving a response time of less than one second.

Nevertheless, the technological maturity of these systems requires further development. Their performance stability is vulnerable to interference in clinical settings; specifically, recognition accuracy can drop to a range of 68% to 75% [75] when multiple individuals are present simultaneously. This limitation currently hinders their broader adoption in large-scale applications.

3.3.4 IoT-integrated AI systems (6 studies, 13.3%).

This category represents the most comprehensive management solution, establishing a closed-loop system that spans from monitoring to intervention by integrating intelligent hand sanitizer dispensers with multi-sensor networks and communication technologies such as Long Range Wide Area Network (LoRaWAN). These systems not only automate the generation of compliance rate reports for HCWs [74] but also employ both mandatory and guided measures, including access control linkages. Some integrated approaches have reported compliance improvements of 30% to 45% [49], suggesting possible behavioral benefits under specific implementation conditions, although broader real-world validation remains limited.

In a burn center study, deployment of such a system combined with weekly individualized feedback increased compliance from 58.5% to 80.5% [73], suggesting that implementation strategies may enhance the impact of these systems under specific deployment conditions. However, deployment costs may still represent an important barrier to widespread hospital-scale implementation.

4.1.1 Overcoming limitations of traditional monitoring methods.

The integration of AI has the potential to address many of the inherent constraints of conventional hand hygiene monitoring. Compared with human observation, AI-based systems enable continuous and objective monitoring while minimizing Hawthorne-related distortion. Unlike human auditors whose physical presence triggers immediate behavioral changes, AI systems operate in the background. Nevertheless, AI systems may still be subject to a residual Hawthorne effect, because awareness of monitoring itself can influence behavior even in the absence of a human observer [82]. Consequently, AI systems may mitigate the Hawthorne effect rather than completely negate the psychological impact of surveillance. For instance, a study conducted in a burn center [73] demonstrated that AI-based automatic monitoring reported compliance rates approximately 12 percentage points lower than those obtained through manual observation, providing a more accurate representation of real-world practice.

Compared to indirect estimation methods relying on sanitizer consumption, AI systems can precisely identify “who, when, and where” missed hand hygiene opportunities, as demonstrated by a LoRaWAN-based system [74]. This capability allows infection control personnel to target interventions to specific individuals and contexts, shifting hand hygiene surveillance from retrospective review toward more proactive management.

4.1.2 Enabling precision in hand hygiene management.

AI-based monitoring systems can detect hand hygiene events while also assessing procedural quality and standardization, addressing a key limitation of traditional methods that emphasize frequency over quality. For instance, one intelligent guidance system [72] identifies whether HCWs omit essential steps such as “rubbing fingertips” or “cleaning wrists.” Similarly, deep learning-based quality assessment research [51] has demonstrated the ability to generate procedural quality scores, improving the proportion of HCWs meeting both duration and quality standards from 45% to 72%.

Systems with real-time alerts can also notify users immediately when hand hygiene is missed. Some studies suggest higher compliance with real-time feedback than with delayed feedback. This shifts AI-based monitoring beyond documentation toward more active behavioral guidance.

4.1.3 Tailoring surveillance to diverse clinical settings.

Based on the currently available evidence, different AI technical pathways may be better aligned with different clinical scenarios according to infection control needs, privacy requirements, mobility patterns, and infrastructure conditions. Fixed clinical areas such as operating rooms and ICUs may be more amenable to computer vision systems. One investigation [76] reported a hand hygiene event detection rate of 93.7% in a multi-bed ICU, suggesting potential utility in this specific setting for capturing transient behaviors among staff frequently entering and exiting patient zones.

For dynamic settings including outpatient and emergency departments, wearable sensor systems may offer a more adaptable option. In privacy-sensitive environments such as general wards or isolation rooms, radar/radio frequency-based systems [75] may offer a non-visual monitoring option based on limited currently available evidence. At the institutional level, IoT-integrated systems may facilitate more centralized monitoring and data aggregation under infrastructure-supported implementation conditions. For example, a LoRaWAN-based system [74] covering 100,000 square meters enabled unified data aggregation across multiple wards, providing continuous quantitative data for infection control decision-making.

4.1.4 Interpreting technical performance in relation to clinical and implementation outcomes.

Although many included studies reported favorable technical metrics such as accuracy, sensitivity, or F1 score, these measures should not be interpreted as direct indicators of clinical effectiveness. In the context of hand hygiene monitoring, high algorithmic performance primarily demonstrates that a system can detect or classify predefined behaviors under specific validation conditions. By itself, it does not establish that the system reduces healthcare-associated infection burden or improves care quality. Nor does it demonstrate sustained behavior change over time or acceptability and feasibility for routine use by HCWs [83]. In the present review, most studies remained focused on technical development, controlled validation, or limited pilot deployment, whereas only a small number examined broader implementation issues such as workflow integration, user acceptability, or sustained compliance improvement [84,85]. Moreover, outcomes that are most meaningful for infection prevention practice, including long-term maintenance of behavioral change, integration with infection surveillance data, cost-effectiveness, and organizational impact, were rarely evaluated. Therefore, the current evidence base mainly supports technical feasibility and early implementation potential rather than established patient-centered or system-level benefit.

Nor is there sufficient evidence at present to show that AI-based monitoring provides superior infection prevention benefit, cost-effectiveness, or sustainability compared with well-implemented multimodal hand hygiene improvement programs [86]. Future research should more explicitly connect AI-based monitoring with pragmatic infection control outcomes, including longitudinal compliance trajectories, staff experience, unintended consequences, and linkage with infection surveillance data. In particular, AI-based nudge strategies may be useful for improving compliance, but their sustained effectiveness and contextual suitability still require evaluation in real-world clinical environments [85].

In addition, implementation studies should consider the possibility of automation bias, whereby infection control teams or frontline users may over-rely on imperfect AI outputs, especially when system benefits are perceived as high [87]. Reported accuracy and related performance measures should be interpreted in relation to validation context and evidence maturity. Metrics derived from technical development or laboratory validation studies mainly reflect performance under constrained conditions, whereas findings from pilot clinical evaluations suggest early implementation potential rather than robust evidence of generalizability or clinical readiness.

4.2.1 Prioritizing scenario adaptability.

In scenarios characterized by varying risk levels, the selection of AI systems should be interpreted in light of contextual adaptability, based on the currently available evidence. For high-risk, stationary environments such as ICUs or operating rooms, computer vision systems may be more suitable, utilizing multi-angle camera coverage to mitigate occlusion effects. In highly mobile areas, such as outpatient clinics or emergency departments, lightweight wearable devices may be advantageous to balance mobility with recognition accuracy. In privacy-sensitive zones such as wards and treatment areas, radar/RF systems or privacy-preserving vision systems [28] may support monitoring while reducing information leakage risks.

4.2.2 Balancing cost, effectiveness, and implementation burden.

Economic considerations remain insufficiently addressed in the current evidence base. Although some AI-based systems may reduce manual auditing workload or improve the timeliness of feedback, few studies formally evaluate whether these technologies are cost-effective relative to established multimodal hand hygiene improvement strategies [25,86]. The available evidence suggests that electronic or automated monitoring may offer operational advantages, but these benefits must be weighed against hardware costs, installation requirements, calibration, maintenance, staff training, and information technology support [25]. Accordingly, the practical value of AI systems should not be judged only by algorithmic performance or short-term compliance gains, but by whether they generate meaningful infection prevention benefit at an acceptable implementation cost. This issue is particularly important for low- and middle-income settings, where competing resource priorities may make high-cost monitoring platforms difficult to justify unless they demonstrably outperform lower-cost multimodal interventions [86,88].

4.2.3 Compatibility with existing systems.

Compatibility of AI systems with existing hospital infrastructure and information systems may lower implementation barriers. If a closed-circuit television system is already operational [76], algorithmic modules may be added to existing camera systems, enabling low-cost upgrades. Similarly, if smart sanitizer dispensers are deployed [74], adding sensor modules can facilitate interoperability with the AI system, thereby obviating the need for redundant equipment procurement.

4.3.1 Technological challenges and optimization strategies.

Despite their advantages in consistency and objectivity, the clinical implementation of AI systems remains constrained by several technical and operational barriers. Firstly, environmental adaptability and model generalizability remain limited. Variations in lighting conditions, camera angles, and individual motion patterns among HCWs can all compromise recognition accuracy. Future research should strengthen cross-institutional data sharing, federated learning, and standardized motion databases to improve robustness across clinical settings. For example, multiple hospitals could collaborate to develop a common hand hygiene motion database based on standardized annotation protocols, shared definitions of hand hygiene opportunities and procedural steps, and consistent metadata on clinical setting, camera position, and participant role. Under a federated learning framework, each institution could retain raw video or sensor data locally while sharing only model parameters or encrypted updates [89,90]. This would allow algorithms to be trained across heterogeneous sites without centralizing sensitive behavioral data. Such an approach may improve cross-site generalizability while also supporting privacy protection and governance compliance.

4.3.2 Practical implementation, human factors, and organizational readiness.

Beyond technical performance, the findings of this review suggest that successful implementation of AI-based hand hygiene monitoring systems is likely to depend on workflow fit, user experience, and organizational conditions. As summarized in Table 5, included systems differed substantially in how they were embedded into practice, ranging from passive background surveillance and retrospective review to real-time prompting, sink-based guided training, wearable reminders, and hybrid monitoring-feedback architectures [30,35,45,73]. These different integration modes are likely to have distinct implications for workflow disruption, behavioral immediacy, and operational burden.

Human factors were addressed unevenly across the literature. A subset of studies reported privacy-preserving designs, anonymized sensing, interface preferences, or practical issues related to wearability and device placement [28,46,75,80]. However, most studies did not systematically evaluate user acceptance, trust, perceived fairness, alert burden, or usability in routine clinical work. This gap is important because even technically accurate systems may fail if they are experienced as intrusive, disruptive, difficult to use, or poorly aligned with frontline workflow.

Organizational readiness was also seldom examined directly. Many systems required substantial local infrastructure, including multiple cameras, sink-mounted devices, wearable tags, wireless communication systems, and continuous calibration or maintenance [49,60,74,76]. These findings suggest that implementation should not be understood solely as a technical problem, but also as an organizational one involving governance, training, workflow redesign, IT support, and long-term operational sustainability.

Taken together, these findings indicate that future evaluations should move beyond technical feasibility alone and examine whether AI-based hand hygiene systems can be integrated into routine practice in ways that are acceptable to users, operationally sustainable, and responsive to the realities of different healthcare settings.

These challenges may be especially pronounced in low-resource settings. AI-based hand hygiene monitoring systems often presuppose several enabling conditions [88]. These include reliable electricity supply, stable network connectivity, routine equipment maintenance, procurement pathways for replacement parts, locally available technical support, and sufficient digital literacy among end users and managers. In settings where infection prevention teams are already understaffed and essential hand hygiene resources remain inconsistently available, introducing complex AI infrastructures may be less feasible than strengthening basic multimodal hand hygiene programs [86,88]. Therefore, the absence of evidence from low-income countries should not be interpreted simply as a research gap, but also as a signal that the infrastructural and workforce assumptions underlying many AI systems may not yet be transferable across settings. Future research should therefore incorporate implementation-science approaches to evaluate adoption, acceptability, feasibility, fidelity, sustainability, and context-specific barriers alongside technical performance.

4.3.3 Ethical and institutional considerations of AI-based monitoring.

In addition, the current evidence base provides only limited insight into algorithmic bias and differential performance across users or contexts. Across the included studies, formal subgroup fairness analysis was rarely reported. Nevertheless, several studies identified subject-dependent or context-dependent variation, including weaker performance of generalized compared with personalized models, reduced accuracy under different camera viewpoints or occlusion conditions, and lower robustness in uncontrolled real-world settings [42,54,76]. These findings suggest that apparently strong aggregate performance metrics may obscure uneven system behavior across staff roles, environments, or user characteristics.

These systems may collect or infer worker-linked behavioral data through video, location, wearable, or event-log signals. Therefore, the ethical issues extend beyond privacy alone and should also be considered in relation to data governance and applicable legal frameworks. In settings where staff can be directly or indirectly identified, such data may fall within the scope of data protection laws such as the General Data Protection Regulation (GDPR). Where monitoring data are linked with identifiable health information handled by covered entities or business associates, obligations under the Health Insurance Portability and Accountability Act (HIPAA) may also become relevant. Accordingly, future implementations should clearly specify what data are collected, for what purpose, and who can access them [91–93]. They should also define how long the data are retained, whether they are de-identified, and whether they may be used only for quality improvement rather than punitive performance management. Broader healthcare AI literature similarly notes that acceptable overall model performance may coexist with biased or uneven behavior driven by data imbalance, measurement bias, and context-specific model development [94]. Future studies should therefore report subgroup-stratified performance, for example across staff roles, care settings, camera viewpoints, and user-dependent versus user-independent workflows, using measures such as sensitivity, specificity, false-alarm rates, and calibration rather than overall accuracy alone [95].

The implications of surveillance-based monitoring systems for workplace culture also warrant closer attention. Although these technologies are often framed as tools for supportive quality improvement, continuous monitoring may also be perceived as managerial surveillance if governance boundaries are unclear or if data are repurposed for punitive evaluation. In practice, this implies the need for explicit governance safeguards. These may include staff consultation before deployment, role-based access control, audit trails, clear retention limits, and restrictions on the secondary use of monitoring data. Such dynamics may erode trust, reduce psychological safety, and weaken healthcare worker engagement with infection prevention efforts. Qualitative research on video-based hand hygiene monitoring has similarly suggested that HCWs’ responses to surveillance-oriented systems are context-dependent and may be shaped by concerns about punitive consequences, confidentiality, data security, patient privacy, and the way feedback is delivered [96].

Automation bias represents another underexamined implementation risk. Infection prevention teams or local managers may over-rely on AI-generated dashboards, compliance summaries, or alerts, particularly when outputs are framed as objective or precise. However, imperfect classification, context-insensitive outputs, and limited generalizability mean that these systems should be interpreted as decision-support tools rather than substitutes for professional judgment, contextual review, or local infection prevention expertise. This concern has also been raised more broadly in clinical AI implementation, where assistive systems may foster misplaced confidence and reduce critical oversight [97].

Finally, the effects of feedback mechanisms are unlikely to be uniform across institutions. Real-time reminders, individualized alerts, or team-level reports may function differently depending on local safety climate, leadership framing, and staff trust in monitoring processes. In supportive, learning-oriented environments, feedback may reinforce improvement; in punitive or low-trust settings, the same feedback mechanisms may provoke resistance, disengagement, or performative compliance. Recent hand hygiene research agendas have specifically identified institutional safety climate as a key determinant and priority area for understanding how feedback and improvement strategies translate into sustained hand hygiene performance [10].

For this reason, future implementation research should evaluate not only technical validity, but also fairness, accountability, organizational context, and the unintended consequences of AI-based monitoring in routine clinical environments.