Generating Explanations to understand Fatigue in Runners Using Time Series Data from Wearable Sensors

Running while fatigued poses an increased risk of injury. Wearable sensors can be used to capture the running kinematics or running pattern as time series signals. The changes that happen in the running pattern due to fatigue, although prominent enough to increase the risk of injury, are generally only seen as subtle differences in the signal itself and hence are difficult to differentiate using purely visual inspection. In this paper, we introduce a time series dataset of motion capture data from runners before and after a fatiguing intervention. The total dataset consists of more than 5500 instances and was collected from 19 participants. The evaluation presented in this paper first looks at the effectiveness of a data aggregation technique called time series barycenters which is shown to improve classification performance. We evaluate and compare a set of classifiers and explanation methods for this problem, and select the most informative classifier and explanation for this dataset. We then present feedback from a domain expert on the insights offered by the the explanations.