Using Machine Learning to Support Physical Exercise, Particularly for Marathon Runners

The rapid adoption of wearable sensors means that massive amounts of data are being collected everyday from people as they exercise. There is an opportunity to use this data with machine learning and recommender systems to help people exercise more effectively. This thesis harnesses such data from 400,000 Strava users as they train for and compete in marathons to provide three beneficial supports: marathon performance prediction, injury risk assessment, and personalised training recommendations. Accurate performance prediction is imperative for effective training and race day planning, and the present work extends previous work by developing a novel case based reasoning (CBR) system for marathon performance prediction that combines both previous races and training-related features, reaching error rates as low as 7.4% for males and 6.2% for females. Since running is a high-impact sport and is associated with a high risk of injury, providing runners with an understanding of their risk is crucial. Here we developed a CBR model to classify runners as likely to sustain an injury or not, and developed a risk score that was highly correlated with the actual prevalence of injuries, using training disruptions as a proxy for injury. We also provide a novel large scale data analysis of the frequency of training disruptions and their cost on marathon performance. Finally, since many runners are recreational runners, they often lack the expertise to tailor their training plans effectively. Hence, we developed a novel recommender system which leverages CBR and ideas from prefactual reasoning to provide runners with personalised training recommendations for how they can adjust their training to meet their goal. We did this in two ways, offering an entire set of training week features, and providing a more targeted recommendation based on a few features. Our findings indicated that when runners sought an improvement of 5-10\% we could make recommendations for most runners, and these typically required 3-5 training adjustments. These supports were evaluated offline and future work would involve a live user trial to determine the efficacy of the predictions, risk assessment, and recommendations, and how runners respond to this information. The methods described in this thesis can be translated to different running distances and indeed to different endurance sports, for example, cycling or swimming.