A cost-effective embeddable algorithm for accelerometer-based fall detector

Falls, often causing injuries in older individuals, involve an unintentional descent to a lower level, like the ground. With the aging global population, addressing fall risks is crucial. Besides falls, delayed medical services post-fall may cause secondary complications. Automated fall detection (FD) systems can promptly identify falls and alert responders. Among automated FD systems, wearable sensor-based ones seem most viable for widespread use. These effectively distinguish falls from daily activities using machine learning techniques. However, their high computational complexity increases power consumption, requires powerful processors, and consequently raises costs. This underscores the need for affordable, embeddable algorithms. Developing highly accurate embeddable algorithms with manageable computational costs remains a current research challenge. This study introduces an algorithm tailored specifically for embedded systems, focusing on ease of implementation and reliance solely on accelerometer data. Empowered by a novel feature, the algorithm integrates thresholding and machine learning techniques, resulting in low computational complexity while maintaining highly effective FD capabilities. Evaluations of the algorithm on comprehensive public fall datasets, KFall and SisFall, demonstrate accuracies exceeding 99% and 97%, respectively. Furthermore, validation on real-world fall events from the FARSEEING dataset yielded an accuracy of 77.3%. Additionally, the proposed algorithm underwent real-time offline analysis on a low-power embedded device. The computational complexity of the proposed method is assessed by comparing it with another low-cost algorithm. Comparative evaluations against a low-cost algorithm, deep learning-based methods, and findings from the literature emphasize the superior performance and cost-effectiveness of this algorithm. Furthermore, the algorithm’s robustness is confirmed through testing at various sampling frequencies, highlighting its ability to achieve successful FD independent of sampling frequency.