A Comparative Study of Deep Learning Architectures for Multi-Horizon Behavioural Forecasting for Mobile Health

This study benchmarks six deep learning architectures, including PatchTST, TCN, MLP, and Transformer, alongside two zero-shot Foundation Models (FM) on three public datasets for multi-horizon behavioral forecasting in mobile health. Key findings reveal that while no single architecture dominates, PatchTST performs best among trained models, and the FM TimesFM excels in low-data scenarios. The research highlights the effectiveness of participant-level fine-tuning, which significantly reduces RMSE, offering valuable insights for practitioners in selecting models and implementing personalization strategies in health-related forecasting applications.