Wearables Group
We develop reproducible methods to analyse time-series wearable sensor data in very large population health studies to better understand the causes and consequences of disease.
Smartphones and wearable devices provide a major opportunity to transform our understanding of the mechanisms, determinants, and consequences of diseases. For example, around 9 in 10 people own a smartphone in the United Kingdom, while one-fifth of US adults own wearable technologies. This high level of device ownership means that many people could contribute to health research from the comfort of their home by offering small amounts of time to share data and help address health-related questions that matter to them.
Our team has played a key role in the collection of wearable sensor data in over 150,000 research participants across the UK and China. This has led to important new discoveries of: new genetic variants for sleep and activity; small amounts of vigorous non-exercise physical activity being associated with substantially lower mortality; and no apparent upper threshold to the benefits of physical activity with respect to cardiovascular disease risk.
We produce open papers, code, and data where our work can be broadly categorised into three major areas:
- Self-supervised machine learning: We develop foundational methods to measure physical activity and sleep status from wearables data in large-scale biobanks that are actively used by researchers worldwide to demonstrate new associations with cardiovascular disease, cancer, depression, mood disorders, and others. These methods generally involve reproducible time-series approaches to robustly identify both behavioural (e.g. walking) and disease-specific (e.g. heart failure) phenotypes. We are actively working on new sensing modalities including cameras, electrocardiograms, and photoplethysmograms.
- Epidemiology of wearable measurements: We develop methods to account for the unique nature of wearable measurements in relation to health association and risk prediction analyses. This has helped us show a clear inverse association between the amount of total physical activity and cardiovascular disease incidence, with no threshold of effect at low or high levels. Our methods generally involve reproducible compositional data analysis to account for the relative time spent in different behaviours over the 24-hour day. We are actively working to expand our work to include large-scale wearable epidemiology studies across a diverse range of international cohorts.
- Genomics of wearable measurements: The combination of wearables with genomics is important to identify and prioritise which wearable measurements are causally associated with disease and to discover new target mechanisms for disease. This has helped us discover the first genetic variants associated with machine-learned sensor phenotypes, showing the first genetic evidence that physical activity might causally lower blood pressure. We are actively working to expand this work to include more international datasets and more detailed genomic information.
Our team
-
Aiden Doherty
Professor of Biomedical Informatics
-
Stefan van Duijvenboden
Researcher in Health Data Science
-
Laura Brocklebank
Senior Researcher
-
Shing Chan
Postdoctoral Research Scientist in Statistical Machine Learning for Cardiovascular Medicine
-
Ben Maylor
Post doctoral researcher
-
Chodziwadziwa Kabudula
Senior Researcher in Wearable Sensors
-
Tatiana Plekhanova
Postdoctoral Researcher
-
Hang Yuan
Early Career Research Fellow
-
Aidan Acquah
Postdoctoral Researcher in Wearable Sensors
-
Vic O'Callaghan
Research Project Assistant
-
Lynne Morley-Johnson
Personal Assistant
-
Alaina Shreves
DPhil student
-
Charilaos Zisou
DPhil student
-
Evelynne Fulda
DPhil student (NIH Oxford-Cambridge Scholars Programme)
-
Yiwei Yue
DPhil Student
-
Anna Bator
CDT student
resources
Data
Capture-24
A human activity recognition dataset with raw accelerometer data and accompanying annotations inferred from wearable cameras.
UK Biobank
Derived physical activity and sleep variables in 100,000 UK Biobank participants.
The main manuscripts associated with this:
China kadoorie biobank
Derived physical activity and sleep variables in 20,000 China Kadoorie Biobank participants.
Code
Please note that non-academics need to purchase a license. Please contact Karl Dickinson for further information
Stepcount
An internally and externally validated software tool to infer step counts from wrist-worn accelerometers.
Actipy
A software tool to ingest multiple activity files into python dataframes with associated calibration, filtering, etc.
Sleepnet
An internally and externally validated (versus polysomnography in ~1700 people) software tool to infer sleep metrics from wrist-worn accelerometers.
Biobank Accelerometer analysis
Software tool used to generate activity variables for the UK Biobank.
SSL Feature Learning
A software tool to extract new movement features (over traditional time and frequency domain features) trained using data from 100,000 UK Biobank participants.
Epicoda
R package designed to support epidemiological analyses using compositional exposure variables.
Training
Join us at Jesus College Oxford for an immersive week-long residential post-graduate short course on machine learning of wearables in large scale biomedical studies. This course aims to connect post-graduate and post-doctoral researchers from academia and industry. Our friendly tutors, internationally recognised for their scientific expertise, will offer specialist instruction and hands-on practicals across five broad areas at the intersection of wearables, machine learning, and health data science: real world validation, time-series machine learning, and hands-on epidemiological analysis in a large dataset. The course is aimed at trainee scientists actively engaged in health data science research, who wish to expand their knowledge of concepts and techniques.