Developing Transferable Deep Models for Mobile Health

Thumbnail Image
Nagesh, Supriya
Rehg, James M.
Inan, Omer T.
Associated Organization(s)
Supplementary to
Human behavior is one of the key facets of health. A major portion of healthcare spending in the US is attributed to chronic diseases, which are linked to behavioral risk factors such as smoking, drinking, unhealthy eating. Mobile devices that are integrated into people's everyday lives make it possible for us to get a closer look into behavior. Two of the most commonly used sensing modalities include Ecological Momentary Assessments (EMAs): surveys about mental states, environment, and other factors, and wearable sensors that are used to capture high frequency contextual and physiological signals. One of the main visions of mobile health (mHealth) is sensor-based behavior modification. Contextual data collected from participants is typically used to train a risk prediction model for adverse events such as smoking, which can then be used to inform intervention design. However, there are several choices in an mHealth study such as the demographics of the participants in the study, the type of sensors used, the questions included in the EMA. This results in two technical challenges to using machine learning models effectively across mHealth studies. The first is the problem of domain shift where the data distribution varies across studies. This would result in models trained on one study to have sub-optimal performance on a different study. Domain shift is common in wearable sensor data since there are several sources of variability such as sensor design, the placement of the sensor on the body, demographics of the users, etc. The second challenge is that of covariate-space shift where the input-space changes across datasets. This is common across EMA datasets since questions can vary based on the study. This thesis studies the problem of covariate-space shift and domain shift in mHealth data. First, I study the problem of domain shift caused by differences in the sensor type and placement in ECG and PPG signals. I propose a self-supervised learning based domain adaptation method that captures the physiological structure of these signals to improve transfer performance of predictive models. Second, I present a method to find a common input representation irrespective of the fine-grained questions in EMA datasets to overcome the problem of covariate-space shift. The next challenge to the deployment of ML models in health is explainability. I explore the problem of bridging the gap between explainability methods and domain experts and present a method to generate plausible, relevant, and convincing explanations.
Date Issued
Resource Type
Resource Subtype
Rights Statement
Rights URI