Fall Detection with Motion Sensors: A Hierarchical Attention-based Convolutional Neural Network Approach

Falls are among the most life-threatening events that challenge senior citizens’ independent living. Wearable sensor technologies have emerged as a viable solution for fall detection. However, existing fall detection models either focus on manual feature engineering or lack explainability. To advance the state-of-the-art of wearable sensor-based health management, we follow the computational design science paradigm and develop a deep learning model to detect falls based on wearable sensor data. We propose a Hierarchical Attention-based Convolutional Neural Network (HACNN) to optimize the model effectiveness. We collected two large publicly available datasets to evaluate our fall detection model. We conduct extensive evaluations on our proposed HACNN and discuss a case study to illustrate its advantage and explainability, that could guide future set-ups for fall detection systems. We contribute to the information systems (IS) knowledge base by enabling explainable fall detection for chronic disease management. We also contribute to the design science theory by proposing generalizable design principles in model building. © 2021 Taylor & Francis Group, LLC