Multidomain Lightweight Adaboost for Real-Time Fall Detection on Low-Power BLE Sensors

This paper presents a practical and energy-efficient approach to real-time fall detection using a lightweight, interpretable machine learning model on a resource-constrained wearable device. We propose a multi-domain learning framework combined with feature-space normalization to enhance generalization across subjects and data sources. A public dataset is augmented with data from a smaller cohort using an articulated skeleton model. To further improve robustness, we employ L2-normalized features. Inertial data are collected at 250 Hz using an Arduino Nano 33 BLE, with local threshold-based filtering to reduce power consumption by transmitting only potential fall events. A compact AdaBoostM1 ensemble (50 depth-3 decision trees) trained on both real and skeleton-based data achieved 93% accuracy on a 30% hold-out from the ShimFall&ADL dataset, significantly reducing false positives compared to threshold-only methods without deep learning's computational overhead. This approach can enable interpretable, ultra-low-power, and disposable fall detection systems suitable for elder-care and rehabilitation applications.