PREDICTING RUNNING ENDURANCE PERFORMANCE BY POOLING RUNNING KINEMATICS AND KINETICS USING THE MOTIONMETRIX SYSTEM, PHYSIOLOGICAL AND ANTHROPOMETRIC DATA: MACHINE LEARNING AND DEEP LEARNING APPROACHES

Author(s): RUDOMETKIN, E.1, PERHAM, M.1, POON, E.T.3, HO, I.M.K.4, LAM, S.S.2, LAW, V.K.H.2, LEE, G.2, NGAN, C.K.1, Institution: TECHNOLOGICAL AND HIGHER EDUCATION INSTITUTE OF HONG KONG, Country: HONG KONG, Abstract-ID: 914

INTRODUCTION:
MotionMetrix is a novel markerless system estimating running joint kinematics and kinetics. Due to the efficient assessment processes, predictive analytics using machine learning (ML) is feasible. This study aimed to use the data-driven analytical framework to predict and potentially improve the endurance performance of runners.
METHODS:
A total of 131 trained runners have completed the MotionMetrix test on a motorized treadmill using their race pace for 1 minute. Meanwhile, all of them completed an incremental VO2 max test, and the final running speed was obtained. Features selection was performed that the best set of predictors including the running gait kinematics (e.g. hip flexion), kinetics (predicted ground contact time), physiological parameters (e.g. VO2 max), and anthropometric factors (e.g. body fat), gender and age were pooled to predict the final running speed of the incremental test. The ensemble-and-ranking-based technique was obtained by the average rank of 3 different approaches including the filter method (Maximum Relevance Minimum Redundancy), the wrapper method (Recursive Feature Elimination with Random Forest), and the embedded method (Ridge Regression) were applied to select the best set of relevant and non-redundant features for producing the subsequent ML algorithms. ML regressors including single-based (SVM and ANN), ensemble-based (AdaBoost and XGBoost), and deep learning-based (Feedforward Neural Networks) models (DL) were conducted.
RESULTS:
The DL model outperformed the ML methods by 25% on average in terms of Mean Squared Error, Root Mean Squared Error, and Mean Absolute Error. Based on the calculated average rank, the top 5 important features in predicting running endurance (i.e. maximum speed) are VO2 max, ground contact time, age, right hip sagittal moment (or maximum propulsive torque), and step separation.
CONCLUSION:
Using the MotionMetrix system to collect running kinematics and kinetics data efficiently can facilitate the use of ML for big data and predictive analytics. The novel feature selection method using the ensemble-and-ranking-based technique and the DL model provided decent predictive performance. Coaches can take reference from our feature importance to identify the training focuses to improve the endurance performance of runners. Future studies on predicting other running-related outcomes such as the actual competition results or risk of injuries using this approach are recommended.