Predicting ground reaction forces and center of pressures from kinematic data in crutch gait based on LSTM

Abstract

Crutches are of extensive applications in the field of rehabilitation. Comprehensively analyzing the ground reaction forces (GRFs) on both crutches and feet can evaluate the patients’ walking function recovery. Given more force platforms are needed in clinical evaluation for the crutch gait than the normal gait pattern and the resulting high cost, this research proposes a method to predict both ground and foot GRFs during walking with crutches, using kinematic information from motion capture trials. We collected force and motion data, built a musculoskeletal model in Opensim, and computed joint angles and moments of crutch gait. Different Artificial Neural Networks (ANN), including Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and Long Short-Term Memory network (LSTM) were established to test their predictive ability using Leave-One-Subject-Out(LOSO) cross validation method. LSTM model showed the strongest agreement, with r = 0.961±0.050 and nRMSE=13.8 % in the vertical direction of the left foot. The LSTM model was more accurate than the CNN model and more robust than the MLP model in this component. In average of different directions, LSTM model has r = 0.656±0.362 and nRMSE=30.3 %. Further verification of the prediction was executed by computing joint moments. The LSTM model showed great application prospects in crutch gait GRF analysis.