Machine learning health-relevant daily living activities from wearable sensors and smartphones
NDPH/2019/44
Background
Activities of daily living such as physical activity, sleep duration, and social interactions, have been associated with physical and cognitive health outcomes. However much of this evidence is based on self-reported questionnaire data that are a-priori selected, sporadically measured, imprecise, and prone to measurement error. In an effort to address some of these issues, cohort studies now collect objective measures of movement via wrist-worn accelerometer sensors. However, such studies narrowly focus on just a small number of summary statistics of pre-specified movement types such as sleep duration, overall activity, or walking. Therefore, uncertainty exists on the health relevance of patterns of daily living activities, in terms of how: 1) they should be measured, 2) their interpretation, and 3) their association with disease onset.
A large proportion of the population own smartphones, which might offer the ability to continuously, and objectively, measure markers of health status in patients’ everyday lives. For example, in less than a month Dr Hinds’ GameChanger Study (https://joingamechanger.org) was able to recruit a cohort of over 12,000 people through novel smartphone based active gamified tasks. However, current efforts to evaluate the ability of mobile phones to measure activities of daily living are flawed. They rely on training data from unrealistic laboratory, rather than naturalistic free-living, environments. Therefore, new approaches are needed to unlock the potential of using mobile phones to better understand the health relevance of activities of daily living.
RESEARCH EXPERIENCE, RESEARCH METHODS AND TRAINING
The project will initially develop validation methods that use wearable sensors via:
1) Unsupervised machine learning of accelerometer data to assess movement-related activities and disease specific signatures (e.g. Alzheimer’s)
2) Deep learning and supervised computer vision of wearable camera data to assess social-related activities (using pre-labelled datasets)
The project will then evaluate the utility of mobile phones to identify health-relevant activities of daily living via:
3) Supervised learning of mobile phone data only. This will require the remote collection of accelerometer and camera data for labels in a sizeable random subset of our existing mobile phone cohorts.
4) Traditional epidemiological investigations into the cross-sectional association between sensor-measured activities of daily living and disease outcomes.
fieldwork, secondments, industry placements and training
This project will offer a comprehensive training programme in bioinformatics and wearable health sensors at a research institute with state-of-the-art facilities. We are committed to training and mentoring students to success.
Our training programme will help students develop skills in large scale health data analysis. The required statistical and bioinformatic approaches for this project are well established and the group has access to large wearable sensing datasets (e.g. the GameChanger study of ~12k participants with mobile phones, the UK-Biobank study of ~100k participants with accelerometer data, and the CAPTURE-24 wearable camera dataset).
This project will also benefit from extensive collaborative links with statisticians and epidemiologists, meaning the student will be well placed for future research positions in the biomedical sciences.
prospective candidate
A candidate with a BSc, or ideally MSc, in a discipline with a substantive computational component.