Estimation of ground reaction force features using inertial sensors data in running: accuracy assessment and analysis of robustness in relation to gait events identification of 11 algorithms

The aim of the present dissertation was to estimate the principal features of ground reaction force (GRF) during running using inertial sensors (IMUs). Specifically, the accuracy and the precision of 11 algorithms utilizing IMU signals for the quantification of GRF-related metrics (i.e. first and second peaks, loading rate, average load, impulse, and pattern) have been assessed, comparing their performance to a gold standard, constituted by the Loadsol insoles. Successively, the robustness of the analyzed algorithms was evaluated with respect to gait events identification (i.e. initial contact and foot-off) by shifting their position (from –50ms to 50ms). This analysis was performed to evaluate how estimation errors changed in relation to potential misidentification of the running stance on IMU-derived signals. Algorithms were divided according to type of approach distinguishing in neural networks ones and “other”. The neural network-based approaches turned out to be in general the most reliable ones, giving lower bias and dispersion values (e.g. -6.6N bias and limits of agreement 1084.5N for the first peak with Pogson xynorm method). Furthermore, the metrics showing the highest correlation with the gold standard were the second peak, the average load and the impulse. Sensitivity of GRF- features to events detection position revealed to be dependent on the analyzed metric and to be variable according to the considered algorithm. However, neural network-based approaches resulted to be more sensitive to foot-off position, whilst the others to the initial contact and stance window one. These results provide opportunities for further studies, regarding the adjustment of algorithms’ formulations and the analysis of more parameters such as foot-strike technique or running surface that can have an influence on the performance.