A MOBILE SENSING APP FOR MENTAL HEALTH TO SUPPORT FEDERATED LEARNING

Abstract

Smartphones are used by half of the world population. More than 10,000 applications are targeted at Mental health. Available apps are limited in four major ways: One, most apps are designed for the Android platform, 80% of the apps did not consider studying both iOS and Android users. Two, there is a lack of a comprehensive tool to study multiple mental health issues. Three, although these apps collect privacy-sensitive data, 67% of studies did not take the privacy concerns of the users into account. Finally, there is an overhead in terms of battery, internet, storage, and time in centralized data analysis. To overcome the limitations, in this thesis we present the design, development, and field evaluation of two mobile sensing applications called PROSIT and PROSITLite. PROSIT app passively and unobtrusively collects mobile sensor data and periodically transfers the data to secure servers. The app tracks 23 different sensor data and hence serves as a comprehensive tool to study different mental health issues. The app runs on both iOS and Android platforms, thus, accessible to over 98% of smartphone users. We conducted an online survey to evaluate the users’ comfortability with PROSIT and privacy concerns, the results from 491 participants show that users are comfortable to track all the app features. Perceptions about surveillance, intrusion, and data leakage influence users’ comfortability negatively whereas trust, control, and consent have a positive influence in user comfortability. We conducted a pilot study on 18 participants who used the app for 2 weeks. The results of the Principal Component Analysis and K-Nearest Neighbours classifier show 73% accuracy in distinguishing the patients and non-patients. Further, we propose a Federated Learning (FL) framework for mental health monitoring to overcome the overheads and preserve privacy. To lay a foundation for FL, we developed PROSITLite with an anomaly detection algorithm to detect Depression. Results from the feasibility study show PROSITLite is efficient in overcoming the identified overheads. In the future, we aim to train a robust model, with data from the ongoing study and implement on-device training with full implementation of federated learning.