Estimating Vertical Ground Reaction Force during Running with 3 Inertial Measurement Units

- Quantification of biomechanical load is crucial to gain insights in the mechanisms causing running related injuries. Ground reaction forces (GRF) can give insights into biomechanical loading, however, measuring GRF is restricted to a gait laboratory. Developments in inertial sensor technology make it possible to measure segment accelerations and orientations outside the lab in the runners’ own environment. The main objective of this study is to estimate vertical GRF with three inertial measurement units using a generic algorithm based on Newtons second law. When using Newton’s second law, it is known that the mass distribution per corresponding acceleration and filtering settings of the acceleration signal do have an influence on the estimated force. Therefore, filtering settings and the mass of the segments were optimized in this study. To apply Newton’s second law to the full body, the accelerations and masses of every segment should be known. However, this requires >10 sensors. By minimizing the number of segments to three, a setup is created that is less obtrusive. Twelve rear foot strike (RFS) runners performed nine trials at three different velocities (10, 12 and 14km/h) and three different stride frequencies (low, preferred, high), on a instrumented treadmill. Inertial measurement units were placed at sternum, pelvis, upper legs, tibias and feet. An optimization was performed to find the optimal sensor configuration. The root mean squared error (RMSE) between the estimated GRF and measured GRF was used as loss function in the optimization. As performance measure of the algorithm, RMSE, active peak error and Pearson’s correlation coefficient were used. The setup with sensors on the tibia and pelvis showed the best result, with an average RMSE of 0.179 bodyweight, peak error of 3.6% and Pearson’s correlation coefficient of 0.98. Using leave-one-subject-out cross validation, it is shown that the algorithm is generalizable within the population of RFS runners. Model performance decreases with velocity but increases with stride frequency. The main error of the algorithm is seen in the first 25% of the stance phase, however, the general performance is comparable or better than what is described in current literature.