A novel graph-based hybrid deep learning of cumulative GRU and deeper GCN for recognition of abnormal gait patterns using wearable sensors

In this study, a novel graph-based hybrid deep learning model of cumulative GRU and deeper GCN was proposed to discover the most representative spatial–temporal gait dynamic features in non-Euclidean space based on the graph for limb joint coupling, significantly improving the generalization of detecting abnormal gait with wearable sensors. The cumulative GRU module first captures the most valuable temporal dependence features by cumulating the useful temporal dependency information from each gait-graph joint node. And then deeper GCN module constructed by residual connection and pre-activation technique is adopted to exploit the best spatial dependency features from the cumulated temporal dependency features by stacking deeper graph convolutional layers. These extracted spatial–temporal gait abnormality features significantly contribute to classifying abnormal gait-graph accurately. Our simulated abnormal gaits data acquired by multi-sensors positioned on lower limb joints are used to feasibly evaluate our method. The experimental results showed that our method with five graph convolutional layers could reach the highest accuracy of 99.02% compared with existing models. Recall and Precision could achieve approximately 100% to identify a certain abnormal gait. Normal lower limb locomotion changes in gait abnormality could be precisely measured. Our technique could feasibly explore the most representative spatial–temporal gait dynamic characteristics in gait-graph for discriminating abnormal gait with high-generalization.