Monitoring Mental Health with a Multimodal Sensor System and Low-power Specialized Hardware

Abstract

This work presents a system that utilizes smartphone and wearable data to understand human behavior and facilitate mental health monitoring. Although mental disorders are exceedingly prevalent, their diagnostic methods are much more antiquated than their physical ailment counterparts. Mental health diagnosticians employ subjective surveys, professional observations, and patient recall to capture a patient’s changing behavior, but these approaches are severely limited. Mobile devices introduce a unique opportunity to quantify and unobtrusively record data on human behavior. Predictive classifiers can interpret the data and yield meaningful behavior classifications and predict mental health status. With the inclusion of specialized hardware, a secure and energy-efficient system can be developed to identify worrying behavior, predict mental health metrics, and encourage users to seek medical help. In this work, machine learning models associated with worrisome mental health behaviors are developed with existing datasets, optimized for performance, and ported to hardware accelerators. Energy metrics for the models are estimated and used as design considerations for a realizable end-to-end system.