Estimation of ground reaction force during running using consumer-level wearable insoles and machine learning

Data from NURVV Run, a consumer-level wearable technology product, embedding pressure insoles and inertial transducers, were used as an input into a deep learning model for the estimation of vertical ground reaction forces (vGRF) during running. Force data were collected from an instrumented treadmill during a running protocol of mixed gradients and speeds, serving as the gold standard to evaluate the model accuracy. Mean difference in peak vGRF was 0.36 ± 0.26 BW across participants and mean root mean squared error was 0.27 ± 0.15 BW. Model accuracy varied considerably between participants; it would
be expected that a larger dataset with a greater variety of input variables would improve on this. A future version of this model could allow continual assessment of load accumulation during distance running, helping identify early signs of elevated injury risk.