A Fall Prediction Method for the Elderly Based on Graph Attention Networks

With the intensification of the global aging population phenomenon, incidents of elderly falls and subsequent accidents are becoming more and more common. Currently, deep learning models for elderly fall prediction use the data's temporal information mostly, and lack the ability to learn discriminative data's spatial information from different inertial sensors. In order to get instant prediction of a fall event and promptly triggering subsequent devices such as airbags, a method based on graph attention neural networks is proposed in this manuscript. By transforming the common time-series data types in fall datasets into network topological structures and content information of sensor nodes, the method utilizes the characteristics of graph neural networks to extract data's spatial features for subsequent classification. Experimental results show that our method achieves accuracy of 97.78^%, sensitivity of 96.76^%, and specificity of 98.49% on the SisFall public fall dataset, with a pre-fall response time 300ms. This means that it will have enough time to judge if a fall event is happening or not first and then triggering protective devices such as airbags.