Mobile Machine Learning Predictions of Bilateral Ground Reaction Forces During Variable Speed Walking and Prosthesis Use

Abstract

Emerging methods, particularly in machine learning, are working to make completely untethered, mobile biomechanical analysis a feasible reality. Traditional instrumented gait labs remain as the gold standard, though moving beyond the physical constraints of these spaces provides several benefits. The two primary use cases are the (1) study of outside-the-lab environments and (2) developing advanced, biomechanically informed device controllers. This work focuses on the latter. Existing models can accurately predict ground reaction forces, though they are limited due to sensor set, dataset size, or application. In this work we demonstrate the accurate prediction of 3-axis bilateral ground reaction forces in able-bodied participants walking with a prosthesis bypass (Avg. R2: 0.95) and without (Avg. R2: 0.98) by using a set of pressure sensing insoles and foot IMUs. This work explores both sensor importance via an ablation study, as well as model transferability through walking speed (Avg. R2: 0.97) and subject holdouts (Avg. R2: 0.68). These results further contribute to mobile biomechanical analysis and device control, specifically for individuals utilizing a lower-limb prosthesis.