Estimating joint moments during treadmill running using various consumer based wearable sensor locations

We estimated lower limb sagittal plane joint moments during treadmill running using wearable sensors and different commonly used locations. We compared outcomes from supervised recurrent neural network machine learning (ML) models to criterion values from motion capture and inverse dynamics. The normalised root mean squared error between outcomes from the ML model fed with the entire wearable dataset (pressure insoles and inertial measurement units at the foot, wrist, T10, and sacrum) was 8.9%, 13.5%, and 18.2% for the ankle, knee, and hip joint respectively. Removal of any two upper body sensors did not decrease the accuracy of the estimations. This work is a springboard to providing biomechanical feedback to runners to help improve performance and minimise injury risk.