Mental health disorders are increasing in occurrence. They are the largest cause of disability worldwide and the strongest predictor of suicide. Despite their prevalence, the majority of affected people either never seek support, or receive limited to no support from under resourced health systems. Further, finding the right treatment for a specific person is a time consuming and inefficient process, as most interventions are based on studies that find the best treatment for a "typical" patient, rather than tailoring interventions to the patient's genes, environment and lifestyle. Hence, to increase access and efficiency of mental health care, there is a need to develop digital tools that make medicine more precise by using data-driven insights and predictions to aid diagnosis, monitoring, prevention, and treatment of mental health disorders.

This PhD thesis focuses on developing computational methods and models that use user-generated data from multiple data sources including passively sensed smartphone and wearable sensor data, text messages exchanged between users, and the users' interaction logs with web or mobile apps, to analyze or predict mental health outcomes with the goal of making the diagnosis and treatment of mental health disorders more efficient and precise. The biggest problem in precision mental health care is the curse of dimensionality in terms of the feature space, outcomes, and patients. That is, to tailor diagnosis, prevention and treatment to each individual, we need to collect and analyze enormous amounts of data associated with the person's behaviors, environment, and other in-situ features, the person' outcomes (e.g. multiple morbidities, outcomes related to mental and physical health), and contextual, occupational, demographic, and other confounding variables that can affect the patient's health. The curse of dimensionality when working with such multimodal and multidimensional data lowers the reliability of analysis and modeling, and decreases the interpretability of the findings.

During my PhD, I addressed the curse of dimensionality challenge through 5 studies. My thesis makes the following contributions: (1) Presents a machine learning based feature selection method that mitigates the curse of dimensionality in the feature space by decomposing and iteratively reducing the feature space during feature selection. (2) Demonstrates the generalizability of the approach in detecting depression, change in depression, and loneliness, as well as forecasting these outcomes several weeks in advance. (3) Demonstrates that behavioral changes resulting from the stay-at-home mandates during the pandemic are predictive of health outcomes during the stay-at-home period for patients with multiple sclerosis. (4) Demonstrates how we can categorize supporters or identify patient phenotypes based on multiple co-morbid outcomes, thereby mitigating the curse of dimensionality with respect to multimorbidities. (5) Presents a method that visualizes and identifies support strategies that work best in an online mental health intervention for patients in a specific context or situation. (6) Demonstrates that accounting for the patient's history or behavioral context improves model performance for the longitudinal monitoring of some health outcomes such as fatigue.

This work has the potential to minimize suffering, by enabling early diagnosis and frequent monitoring of health outcomes using passively sensed longitudinal behavioral data. My work also had implications for more effective treatments through personalization, and improving the patients' awareness about their own health and treatment.

162 pages

Thesis Committee:
Anind Dey (Co-Chair, University of Washington)
Mayank Goel (Co-Chair)
Geoff Kaufman
Andrew Campbell (Dartmouth College)
Mary Czerwinski (Microsoft Research)

Brad A. Myers, Head, Human-Computer Interaction Institute
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science homepage

This page maintained by reports@cs.cmu.edu