Feature Subset Selection for Detecting Fatigue in Runners Using Time Series Sensor Data

Time Series data collected from wearable sensors such as Inertial Measurement Units (IMU) are becoming popular for use in classification tasks in the exercise domain. The data from these IMU sensors tend to have multiple channels of data as well as the potential to augment new time series based features. However, this data also tends to have high correlations between the channels which means that often only a small subset of features are required for classification. A challenge in working with human movement data is that there tends to be inter-subject variabilities which makes it challenging to build a generalised model that works across subjects. In this work, the feasibility of generating generalisable feature subsets to predict fatigue in runners using a correlation based feature subset selection approach was investigated. It is shown that personalised classification systems where the feature selection is also tuned to the individual provides the best overall performance.