Performance of Regression-Based Models for Real-Time Estimation of Anterior Ground Reaction Forces during Walking

Diminished limb propulsive forces correlate with increased fall risk and reduced mobility. Gait biofeedback retraining, focusing on anteriorly directed ground reaction forces, holds promise for improving limb propulsive forces. However, the current reliance on bulky and expensive instrumented treadmills restricts its applicability beyond the laboratory. Inertial measurement units (IMUs), cost-effective alternatives to treadmills, have shown potential in offline estimation of ground reaction forces. Nevertheless, real-time estimation of anterior-posterior ground reaction forces (AGRFs) using IMUs remains unexplored. This study assessed the real-time efficacy of regression-based models for AGRF estimation during walking. Ten participants walked at varying speeds, while IMU and force plate data were recorded. Using 75% of the data for training, participant-specific models were generated and tested on the remaining 25% of trials without altering temporal parameters. Two regression models were created: an unweighted model and a weighted model. Model efficacy for braking and propulsion was evaluated using intraclass correlation coefficients, absolute error, minimal detectable change (MDC), and 95% confidence intervals. Model estimates of the AGRF times series were evaluated using R² and normalized root mean squared error (NRMSE) values. While both models aligned well with collected data, they fell short in predicting peak propulsion force, surpassing the MDC. Model estimates of the AGRF time series also generated relatively low R² values and relatively high NRMSE values. This performance was slightly inferior to real-time regression models for vertical ground reaction forces. These findings suggest leveraging deep learning-based approaches which may be better suited for handling time series data.