Elderly Fall Risk Prediction with Plantar Center of Force Using ConvLSTM Algorithm

Elderly people are vulnerable to falls due to the decline of balance ability, resulting in physical injury, so early detection is essential for fall prevention and reduction in the elderly. Biomechanical sensor data can provide valuable insight into fall risk. However, extracting features from raw time series is a tough task for traditional machine learning methods. In this paper, an end-to-end trainable model named ConvLSTM was proposed to assess fall risk, which works directly on raw plantar force data. 85 elderly people (46 High-risk and 39 low-risk) were recruited for this study. A Footscan® system was used to collect force data of the whole plantar area when each subject walked at normal and steady speed. Firstly, we use t-test to verify the differences between the two risk groups. And then we compared the performance of ConvLSTM model to a baseline model called DTW-KNN. Experimental results show that the classification sensitivity, specificity and accuracy of the ConvLSTM model are optimally 93%, 94% and 94% respectively, which outperforms the DTW-KNN model. The successful application of this model can accurately assess the risk of falls in the elderly, thus providing an early warning basis for fall intervention.