MKLS-Net: Multi-Kernel Convolution LSTM and Self-Attention for Fall Detection Based on Wearable Sensors

Fall detection systems are vital for identifying falls and ensuring prompt assistance, reducing the risk of severe injuries. As society progresses and health concerns gain more attention, extensive research has been conducted to mitigate the effects of falls. Integrating these systems with the Internet of Medical Things (IoMT) has significantly advanced healthcare and personal safety. This study proposes MKLS-Net, a deep learning model that combines multi-kernel convolution, long-short term memory (LSTM), and self-attention mechanism. MKLS-Net performs feature extraction through multi-kernel convolution, passing coarse-grained features to fine-grained ones, minimizing information loss, and improving differentiation between confusing activities. Both LSTM and self-attention help in extracting relatively important information from time series data. The MKLS-Net model demonstrates good fall detection performance on three publicly available datasets, MobiAct, SisFall, and UniMib-SHAR, with best recognition accuracy of 99.51%, 99.94%, and 99.40%, respectively. In addition, for more analysis, we test the proposed model in the multi-classification stage, and it also shows a high accuracy rate in the SisFall dataset with an average accuracy of 83.91%.