A Novel Additive Attention-Based MICNN-BiLSTM Model for Fall Detection Using Wearable Inertial Sensors

Fall detection is a critical issue in elderly care, as falls can lead to severe injuries and even fatalities. Early and accurate detection of falls is crucial to mitigate harm and reduce healthcare costs. Recent advances in deep learning have shown promise in enhancing fall detection systems by leveraging complex patterns in wearable sensor data. This paper proposes a novel model combining a multi-input convolutional neural network (MICNN), bidirectional long short-term memory (BiL-STM), and additive attention mechanism. The MICNN extracts detailed features from multiple sensors, such as accelerometers and gyroscopes, while the BiLSTM captures temporal dependencies in the data. The additive attention mechanism improves interpretability by emphasizing sensor signals that are most relevant to fall events. Moreover, the additive attentionbased MICNN-BiLSTM model demonstrates reduced validation times, enhancing its suitability for real-time applications. We evaluated the model on two benchmark datasets, KFall, and MobiAct achieving 99.30% and 98.85% accuracy, respectively, outperforming state-of-the-art methods.