Symmetry index-based gait improvement prediction using a CNN-LSTM-attention framework

Abstract

The shortcomings of existing clinical assessments are addressed by using a deep learning (DL) framework. This study will assess the gait improvement of lower limb fractured patients using the publicly available GaitRec dataset. The dataset contains the vertical ground reaction force (GRF) data used to compute biomechanical kinetic features. This framework considers hip, knee, ankle, and calcaneus fractures as lower limb fracture class, and the symmetry indices computation is done between affected and unaffected limbs, and their temporal changes were examined to analyze and track the rehabilitation progress of individual subject. The framework successfully identifies the reduction in the asymmetry between left and right leg features. The overall gait improvement in patients is computed using initial and final session composite asymmetry index (ASI) values and if it is at least 30% then overall gait is improved, demonstrating its robustness and clinical relevance. The suggested convolutional neural network (CNN)-long short-term memory (LSTM)-attention hybrid model outperformed with regression metrics of root mean square error (RMSE) of 1.09, mean absolute error (MAE) by 0.64, and R2 score of 0.98 and classification metrics of accuracy 97.1%, precision of 97.6%, recall of 96%, and F1-score of 96.8% by capturing both the local patterns and temporal dynamics of the gait data.