Using machine learning to improve physical activity guidelines

The BHF-Turing Cardiovascular Data Science Awards (First Call): Unsupervised learning of physical activity markers and their association with cardiovascular disease (joint funding with The Alan Turing Institute)

Current evidence shows that too little physical activity increases the risk of heart disease. However, the evidence we have is largely based on asking people how much physical activity they do, which is not accurate. Modern studies tend to ask participants to wear devices that more precisely track their physical activity, but the analysis of this data also has limitations. As a result, we do not know exactly how much, how often, or what type of physical activity is best for health. In this study, heart scientists and data scientists will collaborate to identify new ways to measure activity using one of the world’s largest health datasets, UK Biobank. They will develop machine learning methods that automatically learn how to measure physical activity from the data coming out of devices worn by study participants. The accuracy of the machine learning methods will be tested against physical activity information collected in everyday life by cameras, and studies that measure people's heart health. The work could help us better understand the types and patterns of activity that are the most beneficial for health. This knowledge will influence future lifestyle advice and public health programmes that aim to get the UK moving more.