A robust machine learning enabled decomposition of shear ground reaction forces during the double contact phase of walking

Abstract

BackgroundDynamic analyses of walking rely on the 3D ground reaction forces (GRF) under each foot, while only the resultant force of both limbs may be recorded on a single-belt instrumented treadmill or when both feet touch the same force platform. Research questionThis study aims to develop a robust decomposition of the shear GRF to complete the most accurate decomposition of the vertical GRF [8]. MethodsA retrospective study of 374 healthy adults records (age: 22.8 ± 2.6 years, speed: 1.34 ± 0.28 m/s) and of 434 patient records (age: 21.3 ± 17.8 years, speed: 0.64 ± 0.19 m/s) were used in a machine learning process to develop a robust predictive model to decompose the fore-aft GRF. The lateral GRF was decomposed by resolving the equilibrium of transverse moments around the center of pressure. ResultsA predictive linear model of the fore-aft GRF under the back foot every 5% of the double contact phase was obtained from 2 predictors: the total fore-aft GRF and the vertical GRF under the back foot. Each predictor uses a time series of 31 samples before and during the double contact. The model performs accurately in healthy (median[IQR] error of 3.0[2.2–4.1]%) and in clinical gaits (7.7[4.7–13.4]%). The error in lateral GRF decomposition is of 5.7[3.9–10.2]% in healthy gaits and of 12.0[7.2–19.2]% in patients under the back foot and about half of that under the front foot. SignificanceThe decomposition of shear GRFs achieved in this study supports the mechanics of walking. It provides outstanding accuracy in healthy gait and also applies to neurologic and orthopedic disorders. Together with the vertical GRF decomposition [8], this approach for the shear components paves the way for robust single limb GRF determination on a single-belt instrumented treadmill or when both feet touch the same force platform in normal and clinical gait analysis.