Deep learning-based recognition of simulated nursing activities for infection risk assessment

Hospital-Acquired infections (HAIs) represent a critical global health challenge, significantly increasing patient morbidity, mortality, and healthcare expenditures. Nurses, as frontline caregivers, play a critical role in mitigating these risks through proper hand hygiene. This study presents a novel deep learning-based approach for simulated nursing activity recognition, designed to support the assessment of patient exposure risks to HAIs resulting from missed hand hygiene practices. Using data collected in a high-fidelity simulated patient room environment with healthy volunteers, 29 common nursing activities were analyzed and categorized into four classes based on their potential risk levels for spreading infections. To establish a robust baseline for this domain, we implemented and benchmarked a suite of state-of-the-art spatiotemporal architectures, including SlowFast, inflated 3D, MViTv2, and Video Joint Embedding Predictive Architecture. This comparative analysis focused on evaluating their capability to distinguish fine-grained clinical interactions. Among these, the SlowFast model demonstrated superior performance, achieving mean accuracy and F1-score of 90.78% ± 4.93% and 78.19% ± 3.77%, respectively using leave one subject out cross-validation. These findings demonstrate the technical feasibility of using computer vision for risk assessment in clinical environments. Future work will explore multimodal integration and refined labeling to further improve classification performance and risk assessment.