Trustworthy fall detection: A responsible AI approach for real-time video monitoring of the elderly

Falls among older adults represent a serious public health problem, particularly for individuals living alone or in assisted-care environments, where delayed intervention can lead to severe injuries or death. Although numerous fall-detection systems have been proposed, many existing approaches suffer from limitations, including high computational complexity, limited interpretability, reliance on centralized video storage, and reduced generalization across diverse environments. These challenges motivate the development of lightweight, explainable, and trustworthy fall-detection framework for healthcare applications. To address these limitations, this paper proposes a vision-based fall detection framework that integrates MobileNetV2 for efficient spatial feature extraction and a Gated Recurrent Unit (GRU) for temporal motion modeling. In addition, a federated learning strategy is employed to support collaborative training across distributed data sources without requiring centralized collection of raw video recordings, thereby reducing reliance on centralized storage while enabling learning from heterogeneous environments. During deployment, a sliding-window inference strategy is adopted to achieve reliable and low-latency fall detection. Model transparency is further supported through Grad-CAM++ visual explanations, while Fairlearn is used to assess performance fairness across different fall types and user groups. Experimental results on the combined CAUCAFall and Le2i datasets demonstrate that the proposed framework achieves an accuracy of 97.14%, a precision of 99.26%, a recall of 96.75%, and an F1-score of 97.99%. Although the achieved accuracy is 2.46 percentage points lower than the highest reported accuracy (99.60%), the competing method employs a substantially more complex multi-stage architecture. In contrast, the proposed framework adopts a lightweight MobileNetV2–GRU architecture while integrating explainability through Grad-CAM++, fairness evaluation using Fairlearn, and decentralized collaborative learning. Furthermore, the proposed framework achieves an inference throughput of approximately 93 FPS on the experimental platform used in this study, demonstrating high computational efficiency.