Evaluation of an Artificial Intelligence-Based System for Standardized Hand Hygiene Monitoring Among Operating Room Personnel

Study Design and Clinical Setting

This study was conducted in a multi-center clinical environment encompassing the operating rooms and emergency departments of three tertiary hospitals. The primary research sites included standard surgical operating rooms (for elective surgeries), emergency operating rooms (for trauma and urgent procedures), and a simulated surgical training laboratory (for supplementary data collection and model pre-calibration). The clinical setting was selected to reflect real-world conditions, including variable lighting, personnel movement dynamics, and workflow pressures, ensuring the system’s adaptability to high-demand medical environments.

Data Set Characteristics

AI-Based Hand Hygiene Assessment System

The core of the system was an enhanced CNN integrated with multimodal data fusion technology, designed to recognize handwashing steps, monitor compliance, and provide real-time feedback. Key components included video/audio data acquisition, feature extraction, step recognition, and compliance assessment. Detailed technical implementations and corresponding algorithms (eg, network architecture parameters and mathematical formulas for convolution operations) are provided in the Supplementary Materials and Supplemental Figure 1.

Outcome Measures

Comparative Model Selection

To assess the performance of the enhanced CNN model, four comparative algorithms were selected based on prior use in hand hygiene and motion recognition research. A traditional CNN served as the baseline to evaluate the added value of architectural enhancements. Support Vector Machine-Back Propagation and Random Forest-Back Propagation models were included for their established application and robustness in health care motion classification under variable clinical conditions. K-Nearest Neighbor was used as a simple, interpretable reference to highlight the advantages of deep learning for high-dimensional video data. All models were trained and evaluated on the same data set using identical metrics to ensure comparability, with technical details provided in the Supplementary Materials.

Ethics Approval, Consent, and De-identification of Video and Sensitive Data

This study was conducted with approval from the Ethics Committee of the hospital (Approval Number: XJS20242281-C-1, Approval Date: October 6th, 2024). The study was performed in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants. Participants were informed that video and audio data would be collected for hand hygiene monitoring purposes, and that their identity would be protected through the measures described below.

To protect health care personnel privacy, the AI hand hygiene assessment system applied a multi-layer de-identification strategy. All videos underwent automated AI-based facial blurring before storage or training, obscuring faces, name tags, and other identifying features; analyses retained only hand and arm regions, with backgrounds cropped or blurred. Video and audio files were stripped of metadata and assigned random identifiers, with any linkage to personal information stored separately in an encrypted, access-restricted database. Data collection was limited to hand hygiene-related movements and sounds, excluding conversations and clinical procedures. De-identified data were retained for up to 12 months for quality assurance. The system complies with HIPAA and GDPR, balancing effective monitoring with robust privacy protection.

Experimental Data Preparation

The primary data sources consist of HD videos capturing handwashing operations in operating rooms, complemented by audio data. This combination enables the multimodal fusion model to comprehensively assess the standardization and completeness of handwashing practices. The hardware and software configurations used in the experiments are detailed in Supplemental Table 1.

Experimental Results and Analysis

A total of 892 valid hand-scrubbing events were included in the final analysis (denominator for all compliance estimates). They were collected from real clinical settings and controlled simulation scenarios, ensuring diversity across professional roles, skill levels, and environments. Model performance was evaluated by comparing accuracy and recall among five algorithms: the proposed improved CNN, traditional CNN, SVM-BP, RF-BP, and KNN (Figure 1).OPEN IN VIEWER

As shown in Figure 1A, the improved CNN consistently outperformed all comparison models across training iterations. Accuracy increased rapidly from 50.78% to a peak of 91.91% at 480 iterations and remained above 90% thereafter, indicating strong convergence and stability. In contrast, the traditional CNN plateaued at approximately 85%, SVM-BP peaked at 87.12%, RF-BP at 80.81%, and KNN remained below 75%. Similarly, Figure 1B demonstrates that the improved CNN achieved the highest recall throughout training, while SVM-BP and RF-BP showed slower improvements and lower final recall. These results indicate superior accuracy, recall, and robustness of the proposed model.

As shown in Supplemental Figure 2, the improved CNN exhibits superior computational efficiency, with training time decreasing from 20.87 s to 17.79 s by 80 iterations and remaining stable thereafter, indicating rapid convergence. In comparison, the traditional CNN demonstrates longer training times and slower convergence, while the SVM-BP model shows lower initial time but reduced convergence efficiency.

Figure 2 presents the confusion matrix for handwashing step recognition. The model achieves high accuracy across most steps; however, misclassifications occur between visually similar actions, particularly between “back-of-hands rubbing” and “finger interlacing rubbing.” Overall, the results demonstrate strong efficiency and robust classification performance, with limitations mainly related to subtle motion distinctions.OPEN IN VIEWER

The enhanced CNN model achieved a handwashing step recognition rate of 94.57% (95% CI: 92.81%-95.93%), an assessment accuracy of 93.25% (95% CI: 91.30%-94.76%), and a handwashing duration compliance rate of 92.68% (95% CI: 90.64%-94.31%). All estimates were derived using generalized estimating equations (GEEs) with robust standard errors to account for clustering within participants and hospitals (three tertiary centers). In clinical deployment, the AI monitoring system significantly increased overall hand hygiene compliance to 94.2% (95% CI: 92.1%-95.8%) compared with 83.5% (95% CI: 80.6%-85.9%) in the control group (difference: 10.7%, 95% CI: 7.3%-14.1%, P < 0.001). Average handwashing duration was 45.7 seconds (95% CI: 44.3-47.1 s) in the intervention group vs 58.3 seconds (95% CI: 56.5-60.1 s) in the control group (mean difference: −12.6 s, 95% CI: −14.8 to −10.4 s, P < 0.001). Subgroup analyses were performed to evaluate the effects of clinical environment (elective OR vs emergency OR vs simulated lab) on system performance and compliance. Recognition accuracy remained high across all settings: 94.1% (elective OR), 93.6% (emergency OR), and 95.2% (simulated lab), with no statistically significant difference between environments (P for interaction = 0.32). Similarly, compliance rates were consistent across centers and environments (all P > 0.05), indicating the system was robust to variation in lighting, workflow pressure, and clinical setting.

Empirical Analysis

Structured interviews were conducted with 15 healthcare personnel (5 chief surgeons, 7 surgical nurses, and 3 anesthesiologists/assistants) to evaluate the AI-based handwashing system in the operating room. Semi-structured, face-to-face sessions lasting ∼30 minutes focused on usability, feedback effectiveness, user acceptance, and suggestions for improvement. Interviews were recorded, transcribed, and analyzed via open coding to identify key themes (Table 1).

OPEN IN VIEWER

Table 1.

Qualitative Interview Feedback Data Table

Interview participants	Usability score	Feedback accuracy score	Acceptance score	Number of improvement suggestions	Summary of common feedback
Chief surgeon 1	3	4	3	2	Occasional false alarms from the system interfering with operations
Chief surgeon 2	4	5	4	1	Warnings are timely but often lack clarity
Chief surgeon 3	2	3	3	3	System complexity with a steep learning curve
Chief surgeon 4	4	4	5	1	User-friendly system interface, requests addition of multi-language support
Chief surgeon 5	3	4	3	2	Proposes reducing unnecessary repeated reminders
Nurse 1	5	5	4	1	Simple system operation but occasional delays
Nurse 2	4	4	5	1	Overall satisfied with the system, hopes for long-term use
Nurse 3	3	3	3	2	Noticeable delays affecting operations during peak periods
Nurse 4	4	5	5	0	System is stable, helps reduce human oversight
Nurse 5	5	4	4	1	Requests the addition of pre-surgery self-inspection function
Nurse 6	4	5	4	1	System feedback volume slightly high, causing interference
Nurse 7	3	3	2	3	Requires more training to improve proficiency
Anesthesiologist 1	5	5	4	0	Overall experience is good, but occasional misjudgment commands
Anesthesiologist 2	3	4	3	2	System functions need simplification, reduce interference factors
Auxiliary staff	4	3	4	1	Expects integration with other sterilization management systems

A 4-week comparative study was conducted between the AI intervention group and the conventional manual supervision group. The denominator for calculating the compliance rate was the total number of observed handwashing events (intervention group: n = 437; control group: n = 455). Compliance rate was expressed as a percentage with 95% CI and standard deviation (SD) to reflect data variability.

Data were analyzed using GEE with robust standard errors to account for repeated measurements within individuals and clustering within hospitals/operating rooms. The intervention group achieved a higher mean compliance rate of 94.2% (SD 4.8%, 95% CI: 92.1%-95.8%) compared with the control group 83.5% (SD 7.2%, 95% CI: 80.6%-85.9%), with an absolute difference of 10.7% (95% CI: 7.3%-14.1%, P < 0.001). The mean handwashing duration was 45.7 s (SD 5.3 s) in the intervention group and 58.3 s (SD 8.6 s) in the control group (mean difference −12.6 s, 95% CI: −14.8 to −10.4 s, P < 0.001) (Table 2). These results confirm that the AI system significantly improved hand hygiene compliance and standardized duration, after adjusting for clustering effects from repeated observations and multiple centers.

OPEN IN VIEWER

Table 2.

Handwashing Data Comparison

Groups	Compliance rate (%)	Average handwashing duration (seconds)	Warning frequency (times/person/day)	Common non-compliant behaviors
Experimental group	94.2	45.7	1.8	Insufficient thumb cleaning, skipping steps
Control group	83.5	58.3	3.2	Insufficient handwashing time, skipping steps
Chief surgeons (n = 5)	91.7	43.2	2.1	Neglecting fingertip cleaning
Nurses (n = 7)	95.5	46.8	1.5	Improper wrist cleaning
Anesthesiologists and auxiliary staff (n = 3)	88.9	49.5	2.4	Skipping finger interlacing rubbing step

Enhancing Perioperative Infection Control

Surgical site infections remain a leading cause of health care-associated infections, accounting for 31% of all HAIs, and are closely linked to inadequate hand hygiene.^1,2 Traditional manual supervision is limited by subjectivity, resource constraints, and workflow pressures. The AI-based system addresses these gaps, achieving a 94.57% handwashing step recognition rate and 93.25% assessment accuracy, translating to a 10.7% absolute increase in compliance (from 83.5% to 94.2%). For a department performing 50 daily surgeries, this reduces non-compliant events by ∼ 5 per day, lowering pathogen transmission risk. Duration compliance of 92.68% meets the WHO ≥40-second standard, ensuring adequate antiseptic contact.^2,3 Real-time feedback through visual and auditory alerts corrects non-compliant behaviors immediately, unlike post-hoc auditing. Automated logs and weekly reports enable infection control teams to identify recurring issues and target training, providing data-driven optimization beyond the capabilities of manual monitoring.

A key limitation and practical strength of this study need to be emphasized: The control group received routine, active manual monitoring, which is known to induce a temporary “observational effect” (Hawthorne effect) and elevate compliance above real-world unmonitored levels. In most clinical practice settings, dedicated, continuous manual supervision is not feasible due to staffing shortages, cost, and competing priorities. Thus, the true baseline compliance in routine, unsupervised clinical care is almost certainly lower than the 83.5% observed in our control group. This limitation actually strengthens the real-world value of the AI system: in settings with no manual monitoring, the AI system would likely yield an even larger improvement in compliance because it provides consistent, automated oversight without requiring human resources. Real-time feedback corrected non-compliant behaviors immediately, unlike post-hoc auditing. Automated logs and weekly reports enabled infection control teams to identify recurring issues and target training, supporting data-driven quality improvement beyond the capabilities of manual monitoring.

Optimizing Surgical Team Workflow

Operating room workflows are highly time-sensitive, and manual hand hygiene supervision can disrupt efficiency through delays and rework, affecting team coordination. ⁷ The AI system mitigates this tension by standardizing hand hygiene while minimizing interruptions. The experimental group maintained a consistent average handwashing duration of 45.7 seconds, within the WHO-recommended range, compared with 58.3 seconds in the control group, saving ∼12.6 seconds per event. For surgeries requiring 10 handwashing events, this reduces time by ∼ 2 minutes per case, alleviating scheduling pressures.

Aligning With Regulatory and Consistency With Global Guidelines

Health care institutions are required by international standards (eg, WHO, CDC, and The Joint Commission) to document hand hygiene compliance for accreditation and infection control reporting.^3,6 Manual monitoring is often incomplete, subjective, and resource-intensive, creating compliance gaps and potential institutional risk. The AI-based hand hygiene assessment system addresses these challenges by providing automated, encrypted logging of each handwashing event, including step completion, duration, and alerts issued. This enables data-driven quality improvement: for example, identifying that “finger interlacing rubbing” is frequently misperformed can prompt targeted training for staff (Figure 2).

The system incorporates all 10 standardized handwashing steps defined by WHO and CDC guidelines, with additional granularity on execution order and motion amplitude, ensuring full alignment with global standards and enhancing objective assessment. To ensure reproducibility across institutions with differing local protocols, the AI model was designed with a modular, configurable protocol engine that allows infection control teams to define institution-specific steps, duration thresholds, and antiseptic requirements without full model retraining. All visual recognition modules remain consistent regardless of local protocol variations, and only rule-based compliance logic is adjusted per site. This architecture ensures generalizability and reproducibility when deployed to hospitals with different local hand hygiene policies, while maintaining core recognition performance. By combining regulatory compliance with precise, step-specific monitoring, the system allows infection control teams to track trends, hospital administrators to generate transparent reports, and surgical staff to maintain high-quality hand hygiene. Overall, this approach strengthens institutional safety culture, supports guideline adherence, and provides actionable insights for continuous improvement in perioperative infection prevention.

Technical Advantages in Clinical Context

Existing hand hygiene tools, such as infrared sensors or basic motion detectors, only detect “washing vs not washing” and cannot assess step-specific compliance or duration, limiting their utility in operating rooms. ² The proposed AI system combines multimodal data fusion (video, audio, sensors) with an enhanced CNN, achieving a 94.57% recognition rate, outperforming 3D-CNN models (∼85%) in dynamic OR conditions.^13,14 Attention mechanisms and adaptive learning ensure robustness across routine and emergency surgeries, while sensor integration verifies antiseptic use, addressing regulatory requirements often missed by vision-only systems.^4,5

This holistic monitoring—covering both procedural steps and material compliance—reduces the “false sense of compliance” seen with traditional tools and supports consistent hand hygiene. Improved compliance provides a mechanistic rationale for reducing SSI risk; however, long-term follow-up is needed to confirm clinical outcomes. Future randomized trials with 12-24 months’ follow-up should use SSI incidence as the primary endpoint to validate sustained benefits and patient safety impact.

Limitations and Future Directions

Despite its value, the system requires refinement to fully support surgical and infection control teams. Misclassification of similar steps (eg, “back of hands rubbing” vs “finger interlacing rubbing”), though rare (∼5%), may affect compliance assessment; future integration of transformer models with 3D-CNN architectures aims to improve recognition of subtle motions. ¹⁵ Although peak-time latency was minimal and rarely disruptive (average 180 ms, 99th percentile <500 ms), edge computing optimization will further reduce rare delays. Real-time user feedback channels were implemented to capture ongoing experience, and future iterations will integrate these inputs more closely into model adaptation. Occasional feedback latency during peak OR activity can disrupt workflow, addressable via optimized edge computing. Data privacy remains critical, with encrypted storage aligned to medical standards, while decentralized solutions may further enhance protection. Expanding to high-risk units and linking compliance to preventable infections rates via EHR integration could broaden the impact and provide definitive clinical evidence.