It is not known exactly how much or what type of physical activity is best for reducing the risk of heart disease, and how often it should be done. Wearable sensors provide more precise measures of physical activity but the analysis of such data relies on preconceptions of what types of activity should be studied. This project aims to apply machine learning methods to this, and other, data sources to better understand the types and patterns of activity that are beneficial for cardiovascular health.
One of six British Heart Foundation funded projects.
Explaining the science
Current cardiovascular health guidelines on physical activity are limited by a reliance on subjective self-reported evidence. To more precisely understand the association between physical activity and cardiovascular disease, studies now collect objective measures of activity via wrist-worn accelerometers that capture high frequency data (~180 million readings per participant per week).
However, current methods to analyse this data usually focus on just a small number of summary statistics derived from accelerometer recordings or employ supervised machine learning of pre-specified activity types such as walking or sitting. The latter relies on a prior knowledge of the markers of activity that might be most beneficial for health, and requires substantial resources to collect and carefully label data.
This project will therefore be using unsupervised learning to identify new markers of activity from approximately 100,000 UK Biobank participants' data. Unsupervised learning is a type of machine learning where the algorithms are designed to identify groupings of data without knowing in advance what the groups will be. The training data used to teach these algorithms is unlabelled, i.e. it hasn't been cleaned up, organised, or classified previously.
To develop a new team of cardiovascular researchers and data scientists to identify new 'markers', or indicators, of what types of physical activity are most beneficial for cardiovascular health.
Using UK Biobank, one of the world’s largest health datasets, machine learning methods will be developed that can automatically learn new markers of activity from device data. These will be tested against both reference physical activity data collected in everyday life from camera data, and data from studies of people's cardiovascular health.
This work could help build a better understanding of the types and patterns of activity that are beneficial for cardiovascular health.
This research has the potential to generate new knowledge on the prevention of cardiovascular disease, the leading cause of death in the UK and globally. The project team is well placed to further develop a larger programme of research, to enhance knowledge of activity-related risk prediction and evidence-based strategies for the treatment and prevention of cardiovascular disease. For example, this could include informing UK Chief Medical Officer physical activity guidelines.