Federated Mental Wellbeing Assessment Using Smartphone Sensors Under Unreliable Participation

Today’s smartphones are equipped with sensors that can track and collect data about users’ everyday activities, which can then be transformed into behavioural indicators of users’ health and wellbeing. Prior studies were focused on centralised machine learning techniques, which transfers all the data to a central server. With modern smartphones being powerful enough to process data locally on a user’s device, federated learning (FL) has emerged as a promising alternative that addresses privacy concerns inherent in centralised setups. This study explores the feasibility of FL models to predict mental wellbeing in a decentralised setting. We also closely evaluate how FL can be applied in such applications in the wild, i.e., where user participation may be inconsistent due to device limitations or privacy concerns inherent in mental health monitoring. To further alleviate demands on edge clients, we incorporate federated continual learning, allowing for adaptive, timely model updates that enhance robustness in real-world mental health applications. In our experiments, we trained tree-based, fully-connected and recurrent neural networks, comparing each time with the centralised approach and random baselines. We also assess the model’s ability to generalise across different users and adapt to temporal changes, ensuring reliability across diverse real-world contexts. The findings suggested that given the widespread use of such devices, FL holds great potential in mood and depression detection while protecting data privacy. Our continual FL achieves similar performance to standard FL, but with added benefit of faster model updates.