Novel Multi-Layer Ensemble Algorithm for Fall Detection using wrist-worn devices

Unattended falls of the elderly are a severe health hazard. Several efforts have been made to use Machine Learning and Deep Learning (ML/DL), Internet of Things (IoT), and wearables for Fall detection in the elderly. This paper presents a custom wrist-worn device built using Qualcomm Snapdragon 820c to detect falls. We have used Dew computing, where the data is processed on the wearable device. This system reduces the latency, and issues related to network dysconnectivity are solved. To ensure minimum latency and maximum accuracy, we have used our multi-layer ensemble of ML/DL algorithms. We call this ensemble algorithm, which we developed as variable weight ensemble algorithm - D(VWE(D)). This ensemble algorithm runs on the End device, equipped with a medical-grade Inertial measurement unit (IMU) and heart-rate sensors. The data is collected from the sensors. The statistical features are extracted from the sensors; the data is pruned using Shapley values. The deep learning algorithms are part of the ensemble, Convolution neural network (CNN), and Multi-layered perceptron (MLP), which are pruned using a modified version of the MIT lottery ticket hypothesis. We have obtained an accuracy of 98.2% with specificity and precision of 100%; also, an F1 Score of over 97% indicates that the results are outstanding. In this paper, we have analysed the accuracies and latency on 820c while using the ensemble algorithm and also analysed the accuracies and latencies of the individual components in the ensemble algorithms.