Combining Smartphone Inertial Sensors and Machine Learning Algorithms to Estimate Power Variables in Standing Long Jump

Abstract

Standing long jump (SLJ) power is recognized as informative of the ability of lower limbs to exert power. The study aims to provide athletes/coaches with a simple and low-cost estimate of selected SLJ power features. A group of 150 trained young participants was recruited and performed a SLJ task while holding a smartphone, whose inertial sensors were used to collect data. Considering the state-of-the-art in SLJ biomechanics, a set of features was extracted and then selected by Lasso regression and used as inputs to several different optimized machine learning architectures to estimate the SLJ power variables. A Multi-Layer Perceptron Regressor was selected as the best-performing model to estimate total and concentric antero-posterior mean power, with an RMSE of 0.37 W/kg, R2 > 0.70, and test phase homoscedasticity (Kendall’s τ < 0.1) in both cases. Model performance was dependent on the dataset size rather than the participants’ sex. A Multi-Layer Perceptron Regressor was able to also estimate the antero-posterior peak power (RMSE = 2.34 W/kg; R2 = 0.67), although affected by heteroscedasticity. This study proved the feasibility of combining low-cost smartphone sensors and machine learning to automatically and objectively estimate SLJ power variables in ecological settings.