Uncertainty-Driven Anomaly Detection for Psychotic Relapse Using Smartwatches: Forecasting and Multi-Task Learning Fusion

2026-05-13

AI Summary

This study addresses the high variability inherent in real-world smartwatch data by proposing a Transformer-based multimodal framework for early detection of psychiatric relapse. The approach features a dual-path architecture: one path predicts heart rate dynamics and identifies deviations, while the other integrates sleep, physical activity, and heart rate signals through multitask temporal modeling. Both paths employ MLP-based uncertainty estimation to generate daily anomaly scores, which are subsequently combined via a late-fusion strategy to synthesize complementary digital phenotypes. The key innovation lies in the novel integration of prediction-uncertainty-driven anomaly detection with multitask temporal embeddings, alongside a dual-path late-fusion mechanism. Evaluated on the e-Prevention Grand Challenge dataset, the model achieves an 8% relative performance improvement over the competition-winning baseline, demonstrating that multimodal physiological signal fusion significantly enhances the robustness of relapse detection.

Abstract

Digital phenotyping enables continuous passive monitoring of behavior and physiology, offering a promising paradigm for early detection of psychotic relapse. In this work, we develop and systematically study two smartwatch-based frameworks for daily relapse detection. The first forecasts cardiac dynamics and flags deviations between predicted and observed features as indicators of abnormality. The second adopts a multi-task formulation that fuses sleep with motion and cardiac-derived signals, learning time-aware embeddings and predicting measurement timing. Both pipelines use Transformer encoders and output a daily anomaly score, derived from predictive uncertainty estimated via an ensemble of multilayer perceptrons to improve robustness to real-world wearable variability. While each framework independently demonstrates strong predictive power, we show that they capture complementary physiological signatures. Consequently, we propose a late-fusion strategy that synergistically combines the anomaly signals from both architectures into a unified decision score. We benchmark our methodology on the 2nd e-Prevention Grand Challenge dataset, where our fused model achieves a 8% relative improvement over the competition-winning baseline. Our results, supported by extensive ablation studies, suggest that the integration of diverse digital phenotypes, cardiac, motion, and sleep, is essential for the high-fidelity detection of psychotic relapse in real-world settings.

Problem

• psychotic relapse
• anomaly detection
• digital phenotyping
• smartwatches
• uncertainty

Innovation

• uncertainty-driven anomaly detection
• multi-task learning fusion
• digital phenotyping
• Transformer encoder
• predictive uncertainty