Federated Learning Framework for Mobile Sensing Apps in Mental Health

Abstract

Mental health issues are negatively impacting people, the economy and life expectancy. Several mobile applications are developed to aid mental health treatment and mobile sensing applications help remotely monitor patients with mental illness, understand key factors like sleep and exercise, and deliver effective treatment methods. Though new smartphones are increasingly efficient, the majority of mental health applications transfer data to centralized servers for processing. In this paper, we propose a Federated Learning framework for Mental Health Monitoring Systems (MHMS) to preserve user data privacy, reduce network usage and improve performance. To detect depression using the Federated Learning framework we defined a mobile application architecture, and developed two versions of applications that collect three types of sensing information such as location, accelerometer and calls. We defined epochs and developed an anomaly detection algorithm that helps to label local data to train models. We conducted a preliminary study for 6 weeks using two app versions. The results from the study indicate the app implementing the Federated Learning framework is capable of continuously tracking data utilizing less power, storage space and internet data. It also preserved users' privacy. In future, we are planning to implement Federated Learning to run large-scale studies with improved server-side federated averaging methods.