SCH: Personalized Depression Treatment Support by Mobile Sensor Analytics

点击翻译按钮获取中文摘要

The current best practice guidelines for treating depression call for close monitoring of patients, and periodically adjusting treatment as needed. This project will advance personalized depression treatment by developing an innovative system, DepWatch, that leverages mobile health technologies and machine learning tools to provide clinicians objective, accurate, and timely assessment of depression symptoms to assist with their clinical decision making process. Specifically, DepWatch collects sensory data passively from smartphones and wristbands, without any user interaction, and uses simple user-friendly interfaces to collect ecological momentary assessments (EMA), medication adherence and safety related data from patients. The collected data will be fed to machine learning models to be developed in the project to provide weekly assessment of patient symptom levels and predict the trajectory of treatment response over time. The assessment and prediction results are then presented using a graphic interface to clinicians to help them make critical treatment decisions. Our project comprises two studies. Phase I collects sensory data and other data (e.g., clinical data, EMA, tolerability and safety data) from 250 adult participants with unstable depression symptomatology. The data thus collected will be used to develop and validate assessment and prediction models, which will be incorporated into DepWatch system. In Phase II, three clinicians will use DepWatch to support their clinical decision making process; a total of 50 participants under treatment by the three participating clinicians will be recruited for the study. A number of innovative machine learning techniques will be developed. These include a set of new learning formulations to construct matrix-based longitudinal predictive models, and determine the temporal contingency and the most influential features, and deep learning based data imputation methods that can handle both problems of sporadic missing data as well as missing data in an entire view. In addition, multi-task feature learning models and feature selection techniques will be expanded and refined for this challenging setting of large-scale heterogeneous data.