Neural networks can accurately identify individual runners from their foot kinematics, but fail to predict their running performance

Abstract

Athletes and coaches may seek to improve running performance through adjustments to running form. Running form refers to the biomechanical characteristics of a runner’s movement, and can distinguish individual runners as well as groups of runners, such as long-distance and short-distance runners. Yet, in long-distance running it is still unclear whether certain running forms lead to better performance. In this study, we used a neural network to test the extent to which individual running forms, measured from foot kinematics, exist within long-distance runners and whether running forms can predict performance. To accomplish this goal, 119 participants ran on a treadmill at three different speeds and overground at a self-selected sub- maximal speed while we collected data from insole-embedded Inertial Measurement Units (IMUs) mounted in both shoes. Participants reported their personal best 10 km run times. We used these data to train the neural network to identify individual runners from their running data. Then, we trained the same neural network architecture to predict the runners’ performance. With enough data, the neural network was successful in identifying individual runners, but was comparable to a random coin flip (57 % accuracy) in predicting whether an individual runner is slow or fast. We interpret the success of the model to identify runners, but the subsequent failure of the same model to predict running performance as evidence that individual running form measured from foot kinematics contains insufficient information about a runner’s performance.