Continual Learning Strategies for Personalized Mental Well-being Monitoring from Mobile Sensing Data

Automatic prediction of psychological stress using mobile sensing data is critical for developing personalized mental health monitoring tools. Traditional approaches typically train static models offline, limiting their applicability in real-world scenarios characterized by evolving user behaviors and data streams. To address these challenges, we formulate stress prediction as a continual learning (CL) regression task, where each user constitutes a distinct sequential task. Within this framework, we evaluate two adaptation strategies: (1) a replay-based approach, leveraging recurrent neural networks combined with an episodic memory buffer; and (2) a prompt-based method, where task-specific prompts are jointly trained with a transformer backbone and combined with the memory buffer to enable task-aware adaptation across users. Experiments conducted on a filtered subset of the StudentLife dataset, comprising multimodal behavioral and contextual data, demonstrate that prompt-based methods achieve competitive predictive performance. Our results underline the potential of continual learning paradigms to enable scalable and adaptive stress prediction systems.