Multivariate Time Series Classification of Exercise Data

Exercise data captured using Inertial Measurement Units (IMU) can be used to train models to objectively classify kinematic behaviour. These models can help to understand performance, diagnose injury, and monitor the recovery of athletes. Running is one such exercise which has become very popular over the past few decades and this increase in popularity has also led to an increase in the rate of running related musculoskeletal injuries. These injuries frequently occur due to faults in the runners technique which can happen as the runner loses control due to fatigue. Wearable sensors such as IMUs can capture this loss in technique. IMU data contains multiple dimensions which tends to be correlated and generally noisy. Hence the selection of useful features is an important challenge in building these classification models. Research in Multivariate Time Series Classification (MTSC) where multiple time series are associated with a class label has become increasingly popular in the last few years. Most recent advances have been in the development of algorithms for MTSC and state-of-the-art techniques are very well developed with methods such as Rocket, or 1-NNDTW which work very well for MTSC. However, there has been less focus placed on feature selection for Multivariate Time Series (MTS) data. This along with the requirement for selecting useful features from the wearable sensor data motivated the development of a feature subset selection technique unique to MTS data and based on the correlations in the data. Furthermore, an important challenge while applying machine learning techniques to exercise data, is that the model outputs should be meaningful to the clinician, domain expert, or the end user. This thesis investigates the use of multiple techniques to improve the usability of the exercise prediction models including (i) feature subset selection to identify the more useful channels of data, (ii) Barycenters as an aggregation and visualisation technique, and (iii) interpretable time series classification to understand which parts of the time series contribute more to the classification task in the context of exercise evaluation. The feature subset selection technique proposed in this thesis to help identify more useful feature subsets was able to significantly decrease the number of features required whilst maintaining the overall accuracy. Another challenge in this domain is the lack of available datasets which limits the application of such machine learning techniques onto exercise data. In this work, we collect and release a dataset of wearable sensor data that captures the onset of fatigue in runners. This dataset is thoroughly investigated in this thesis. In developing machine learning techniques to work with the wearable sensor data, it is crucial that they are developed in a way that can in the future be extended to a real-life setting. An important aspect of making the techniques usable in the personal sensing domain is that it should be generalisable when new individuals are brought into the picture. In this thesis, an exploration into techniques to take personalised models which are currently the norm in this domain and push them into more globalised solutions is explored. A clustering technique proposed in this thesis showed that it was possible to improve performance up to 20% in a globalised setting through this technique.