• Two-stage model integrates metadata to improve hospital hazard detection. • Reduces false detections by leveraging human expertise on hazard occurrences. • Enhanced cross-location generalization without model retraining. • Enables lower detection thresholds while maintaining high precision. • Various object detectors assessed on realistic indoor hazard scenarios. Maintaining environmental safety in hospitals is critical for preventing injuries and infections, yet manual monitoring is resource-intensive and prone to coverage gaps. This study presents a novel, two-stage AI framework for automated hazard detection. We first evaluated three state-of-the-art architectures—YOLOv8, DETR, and EfficientDet—finding that while all models demonstrated strong in-domain performance, they suffered significant accuracy loss in unseen areas due to domain shift. Among these models, YOLOv8 demonstrated superior baseline robustness. To address this generalization gap, we developed a second, metadata-augmented refinement layer that integrates hospital-specific spatial and temporal hazard information. Uniquely, this approach incorporates expert operational knowledge externally during inference rather than embedding it during training, thereby reducing the need for extensive annotated datasets or frequent model retraining. Experiments conducted in a major tertiary medical center demonstrate that metadata augmentation significantly improves detection reliability in cross-domain scenarios. The system operates effectively at lower objectiveness thresholds, achieving the high recall essential for safety–critical applications while simultaneously improving precision. Overall, the proposed framework offers a scalable and data-efficient solution for real-time environmental risk management. By bridging the gap between visual AI capabilities and the rigorous safety demands of healthcare facilities, it provides a practical pathway for enhancing hazard detection, situational awareness, and operational resilience.
Yahav et al. (Wed,) studied this question.