Towards a Fall Detection System: A Deep Learning and Wearable Sensors Approach for the Internet-of-Healthcare-Things Applications

Abstract

Falls among the elderly are a significant global health concern. Detecting falls accu rately and promptly can greatly reduce injuries and related expenses.At present, fall detection systems that use sensor-based data commonly employ conventional machine learning approaches like Support Vector Machines (SVMs) or advanced deep learning methods such as Convolutional Neural Networks (CNNs) and Recurrent Neural Net works (RNNs). Nevertheless, there exists a necessity to investigate the cutting-edge transformer model for potential applications in this domain. In this research, we pro pose a deep learning model inspired by the Transformer Model by using encoder & attention mechanism. We have used the publicly available K-Fall dataset to evaluate the model using 6D input features i.e., accelerometer & gyroscope measurements. We test our approach on three different fall phases: pre-impact, post-impact, and the fall cycle phase, extracting temporal features from falling signals. For binary classification, we distinguish between Falls and ADLs with remarkable results boasting an average ac curacy 96.75% for pre-impact with inference time of 47ms, 99.96% for post-impact with inference time of 42ms & 99.95% for fall-cycle phase with inference time of 49 ms.These results signify that the model effectively minimizes false negatives, crucial for reliable fall detection.