Federated Reinforcement Learning for Personalised Mental Health Interventions

Abstract

Depression and anxiety are among the leading causes of disability worldwide, placing a heavy burden on individuals, families, and healthcare systems. Although awareness and resources for mental health care have grown, access to effective and personalized treatment remains limited. Digital micro-interventions provide short, targeted, and accessible support, yet their effectiveness depends on tailoring them to individual needs. Machine learning methods, particularly Reinforcement Learning, can deliver such personalization by adapting to user behaviors and responses over time. However, these approaches depend on large and diverse datasets, which are becoming increasingly difficult to access, especially in healthcare, where strict privacy regulations are enforced due to the sensitive nature of the data. This creates a central challenge: how to design adaptive and effective Artificial Intelligence systems for mental health while safeguarding privacy. This dissertation explores Federated Reinforcement Learning as a solution to this challenge. Federated Learning enables models to be trained across distributed data sources without transferring raw patient information, while Reinforcement Learning adapts interventions dynamically through user interactions. The integration of these two methods provides a framework for secure, scalable, and personalized digital mental health support. The proposed system was implemented on the Moodbuster platform and evaluated through simulated experiments and a pilot study with actual participants. Results show that the federated approach achieved performance comparable to that of centralized models while preserving privacy, and the reinforcement learning component supported personalized intervention delivery, contributing to user engagement and effectiveness.