Author(s): BOGAERT, S., DAVIS, J., VANWANSEELE, B., Institution: KU LEUVEN, Country: BELGIUM, Abstract-ID: 1167

INTRODUCTION:
Running poses a high risk of developing running-related injuries [1]. The majority of RRIs are the result of an imbalance between the cumulative musculoskeletal load and a person’s load capacity [2, 3]. Therefore, measuring and monitoring the musculoskeletal load during running is essential for the prevention of RRIs. A general estimate of whole-body biomechanical load can be inferred from ground reaction forces (GRFs). Unfortunately, GRF typically can only be measured in a controlled environment, which hinders its wider applicability. Applying machine-learning algorithms to wearables-collected data enables runners to monitor GRF characteristics in various settings [4], extending beyond the confines of the laboratory. Our study presents and evaluates a machine-learning method to predict contact time, active peak, impact peak, and impulse of the vertical GRF (vGRF) during running from 3D sacral acceleration.
METHODS:
Twenty-seven subjects with varying running experience ran on an instrumented treadmill (Motek Medical BV) at various speeds for 150 seconds at each speed. The instrumented treadmill measured ground reaction forces, and a sensor (Xsens Technologies, Movella) on the lower back captured 3D acceleration. In total, we obtained 50072 steps of which 28078 steps had an impact peak. Subject characteristics, general statistical, and domain-specific features were extracted from the 3D acceleration signal for each step.
We used a Lasso (Least Absolute Shrinkage and Selection Operator) model to predict the characteristics of the vGRF. We partitioned the data into a training (23 subjects) and test (4 subjects) set. To select the hyperparameter settings, we conducted a leave-one-subject-out cross-validation on the training set. We compared the performance of our approach to the methods outlined in the work of Alcantara et al. [4] regarding active peak, contact time, and impulse; Verheul et al. [5] for the impact peak using a single-sensor setting.
RESULTS:
The developed models for predicting active peak, impact peak, impulse, and contact time had a root-mean-squared error of 0.084 body weight (BW), 0.221 BW, 0.005 BW.s, and 0.008 seconds, respectively. Our models outperform the corresponding comparison method from the literature on our data sets by respectively 0.061 BW, 0.054 BW, 0.0032 BW.s, and 0.0006 seconds.
CONCLUSION:
These results indicate the feasibility of our approach for monitoring selected factors associated with running-related injuries outside the laboratory with a single sensor.
REFERENCES:
[1] Videbæk, S. et al. Sport. Med. 45, 1017–1026 (2015)
[2] Hespanhol Junior, L. C. et al. Sport. Med. 47, 367–377 (2017)
[3] Bertelsen, M. L. et al. Scand. J. Med. Sci. Sport. 27, 1170–1180 (2017)
[4] Alcantara, R. S. et al. PeerJ 9, 1–18 (2021)
[5] Verheul, J. et al. J. Sci. Med. Sport 22, 716–722 (2019)