Precision VISSTA: Machine Learning Prediction and Inference for Bring-Your-Own-Device (BYOD) mHealth Data


Precision VISSTA is a bring-your-own-device (BYOD) mobile health (mHealth) patient-powered research study focused on Inflammatory Bowel Diseases (IBDs). Participants report longitudinal survey data on outcomes such as disease activity along with patient-reported outcomes like as sleep disturbance, while also contributing mHealth lifestyle data from various wearable devices and apps (24 types). IBD patients have extremely heterogeneous phenotypes with symptoms that fluctuate. Prior work has suggested an association between increased self-reported physical activity and decreased disease activity , while self-reported sleep disturbance has been associated with increased disease activity . However, the precise nature and quantity of activity and sleep associated with improved outcomes is not well established. Our mHealth dataset contains numerous features describing physical activity and a number of other lifestyle characteristics, allowing for large-scale analysis of the features most associated with IBD disease activity and symptoms. Because of the complex underlying relationships within the data, we considered a number of flexible machine learning (ML) approaches in order to avoid the rigid model structure imposed by most traditional statistical models. We leveraged recent theoretical results on inference for supervised learning ensembles to develop and implement permutation-style hypothesis tests for feature significance on these otherwise “black-box” models. The primary study objectives were: (1) to formally establish the predictive relevance of mHealth features in forming more accurate predictive models than could be obtained with survey data alone, and (2) to infer which specific mHealth lifestyle features are most predictive of outcomes for patients with IBDs.

American Medical Informatics Association (AMIA) Annual Symposium Podium Abstract