We aim to make use of observational data to gain insight into human health, and better prevent and treat disease. Our core research areas are thus methods for causal inference and explanation from complex data, discovery from observational biomedical data, and automated dietary monitoring.
How can we learn how things work when we can only observe, and not experiment on a system? This is the problem faced in many biomedical sciences, and is complicated by the fact that causes are not only composed of individual variables but complex entities like conjunctions (exercise and meals interact to affect blood glucose) and durations (smoking for decades is likelier to cause cancer than a single cigarette), and relationships happen at many timescales (weight loss has a gradual impact on glucose, eating affects it sooner). We develop new algorithms that can automatically extract these complex relationships from large-scale time series data.
We aim to improve human health by helping clinicians and patients make better use of complex medical data. Our methods have been applied to gain insight into stroke recovery using ICU data, risk factors for heart failure using EHR data, and causes of hyperglycemia in people with diabetes wearing sensors during daily life. We have developed new technology along the way, including better methods for handling missing data and approaches for directly modeling the uncertainty in measurements.
Nutrition is a core part of maintaining health and preventing disease, but there is not yet a way to automatically and unobtrusively measure diet in the same way we can measure physical activity. We aim to change that, and in laboratory settings designed to mimic daily life while also providing ground truth, we have shown that it is possible to achieve performance on par with humans — without the human labor of logging meals. In ongoing work, we aim to take this technology out of the lab, and make it practical for use in daily life.