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

2026-06-12

AI Summary

This study addresses the lack of systematic comparisons among multistep forecasting models for wearable-derived behavioral time-series data, particularly regarding cross-population generalization, individual fine-tuning, and long-horizon prediction. It presents the first unified evaluation of six deep learning architectures—including PatchTST and Transformer—alongside two zero-shot foundation models such as TimesFM and statistical baselines, across three public datasets comprising over 800 participants. The evaluation focuses on multistep prediction of daily step count, screen time, and sleep duration, incorporating feature-level fine-tuning and cross-dataset transfer experiments. Results indicate no single architecture consistently dominates; notably, TimesFM achieves zero-shot performance comparable to or better than trained models, while individual fine-tuning substantially reduces RMSE across all behavioral metrics by 16–60%, with the most pronounced gains observed for sleep prediction.

Abstract

Wearable devices and smartphones generate rich behavioural time series that can support proactive health interventions, yet systematic comparisons of modern forecasting architectures for these data are lacking. In particular, it remains unclear how models generalise across populations, how different architectures respond to participant-level fine-tuning and how forecasting accuracy degrades across multi-day horizons. We benchmark six deep learning architectures, two zero-shot Foundation Models (FM) and statistical baselines on three public datasets encompassing over 800 participants, reporting per-feature metrics for step counts, screen time and sleep duration across 1-8 day horizons. We further conduct a per-feature personalisation study across all six architectures and assess FM transferability across dataset sizes and temporal granularities. Our key findings are: (i) no single architecture dominates, PatchTST leads among trained models while the three runners-up (TCN, MLP, Transformer) show no meaningful performance difference; (ii) the FM TimesFM matches or exceeds trained models zero-shot, especially in low-data regimes and (iii) participant-level fine-tuning reduces per-feature RMSE by 16-60\%, with sleep benefiting most and step counts least. These results provide practical guidance on architecture selection, FM applicability and personalisation strategies for mobile health forecasting. To the best of our knowledge, this is the first study to jointly evaluate modern deep learning, FMs and personalisation for multi-horizon behavioural forecasting from wearables.

Problem

• multi-horizon forecasting
• behavioural time series
• wearable devices
• model generalisation
• personalisation

Innovation

• multi-horizon forecasting
• foundation models
• personalisation
• wearable time series
• deep learning benchmarking