Deep learning based ground reaction force estimation for stair walking using kinematic data

Abstract

Complete ground reaction forces (GRFs) are vital for biomechanical analysis. The GRFs are currently measured by force plates. The measurement of GRFs during stair walking is difficult due to the need for instrumented staircases. We trained two bi-lateral long short-term memory (BiLSTM) neural networks to estimate 3D GRFs during stair ascent and stair descent using the whole-body kinematics as the input. The dataset is collected from eighty subjects, including healthy and knee osteoarthritis individuals. We also developed a post-processing algorithm to remove artifacts on GRFs in the swing phase. Our models achieved excellent accuracy compared with the measured GRFs with the correlations of 0.908∼0.991, the root mean squared error (RMSE) of 3.29% and 3.56% body weight (BW) and the normalized RMSE (nRMSE) lower than 5% and 8% for the complete GRFs during stair descent and ascent. Using our models, researchers can estimate 3D GRFs during stair walking without instrumented staircases.