Measurement of lifestyle health behaviours from wearable cameras

点击翻译按钮获取中文摘要

The lifestyle choices that people make have repercussions for their health. For instance, reallocating time away from sedentary behaviour has been associated with lower risk of incident cardiovascular disease. Traditionally, identifying these choices has relied on self-reported data, which suffers from information bias, which has meant that the impacts of lifestyle behaviours are sometimes inaccurately estimated. However, advances in wearable sensors have introduced a more objective avenue for measuring lifestyle behaviours. Recent work has shown that wearable accelerometers are accurate at distinguishing between different physical activity behaviours. However, they still have not shown to be able to capture a range of lifestyle behaviours, for instance, recognising people's social interactions. Wearable cameras provide a promising avenue for identifying a wide range of these behaviours, as well as the context in which they occur. The gold standard measurement of lifestyle behaviour is direct observation; reviewing footage from wearable cameras is often as close as we can get to this gold standard. However, in order for these cameras to be included in the next generation of large-scale population health studies, we need to develop methodology for annotating wearable camera data at scale. Although we can take inspiration from the field of video-based human activity recognition, models developed in this field are typically trained on video clips that are very different to the series of images captured by camera loggers, which capture photos only every 20-30 seconds and are worn for a full day of different activities. This project will involve developing novel computer vision methods that can recognise a range of lifestyle activities from the relatively infrequent images captured by camera loggers.During a rotation project with the OxWearables group, I have shown that it is possible to automatically distinguish between coarse descriptions of physical activity from wearable cameras using techniques from computer vision. A DPhil in this direction would aim to rigorously assess the feasibility of incorporating wearable cameras into large-scale population health studies, and would include the following specific aims:1. Develop unsupervised and semi-supervised techniques for activity recognition based on camera logger data, with the goal of developing a tool that improves the annotation of new camera logger data-sets 2. Review the ethical frameworks for working with wearable camera data and revise them to reflect the issues raised by automated classification of lifestyle behaviours, 3. Develop multi-modal approaches to activity recognition that combine wearable cameras and accelerometers, and determine which types of activities are better recognised using different modalities, 4. Assess generalisation of wearable camera based classifiers across different population groups, environments and wearable devices.In the OxWearables group, there is unique access to wearable camera and accelerometer data-sets for physical activity recognition, and there are several new studies in progress which promise to bring in further data, such as the RADAR-AD project comprising 100 EU citizens, a study of 300 adolescents based in Australia, and a study of 150 older adults based in the UK and China. If successful, the methods proposed here promise to objectively measure lifestyle behaviours in future large-scale population health studies, and thus contribute to our understanding of how these behaviours are associated with major disease outcomes.This project falls within the ESPRC healthcare technologies research area. There are no additional companies or collaborators involved in this project.