Federated Learning for Hierarchical Fall Detection and Human Activity Recognition

In healthcare monitoring, precise fall detection (FD) and human activity recognition (HAR) are paramount, especially for the elderly. This paper presents a federated learning (FL) framework that employs a two-stage hierarchical approach to address these needs. The first stage distinguishes between fall and non-fall events, crucial for minimizing false alarms in sensitive environments such as elderly care facilities. Subsequently, if a fall is detected, the system classifies the type of fall to facilitate appropriate medical responses; if no fall is detected, it classifies the specific activity being performed. This approach enables accurate emergency responses and supports personalized healthcare interventions. With FL, all model training are conducted on local devices using wearable data, including inertial measurement unit (IMU) and physiological signals, without the need to share sensitive data centrally, preserving users’ privacy. This method ensures that each device contributes to a global model, whilst maintaining the confidentiality of individual data inputs. Our experimental evaluations demonstrate generalization capabilities in binary classification for FD and highlight challenges in multi-class scenarios for HAR, demonstrating the need for advanced strategies in handling complex classifications. The source code and experimental evaluations are accessible at https://github.com/SIT-FL/FL-Fall-Detection.