Uncertainty-Driven Anomaly Detection for Psychotic Relapse Using Smartwatches: Forecasting and Multi-Task Learning Fusion
2026-05-13 • Machine Learning
Machine Learning
AI summaryⓘ
The authors developed two smartwatch-based systems to detect early signs of psychotic relapse by monitoring heart, movement, and sleep patterns. One system predicts normal heart behavior and flags unusual changes, while the other combines information from sleep, motion, and heart signals to learn daily patterns. Both use advanced machine learning models called Transformers and measure uncertainty to improve reliability. By combining these two methods, the authors achieved better detection results compared to previous approaches. Their work shows that using multiple types of wearable data together helps spot relapse more accurately in everyday life.
Digital phenotypingPsychotic relapseSmartwatch monitoringCardiac dynamicsSleep patternsMotion sensorsTransformer encoderAnomaly detectionPredictive uncertaintyMultitask learning
Authors
Nikolaos Tsalkitzis, Panagiotis P. Filntisis, Petros Maragos, Niki Efthymiou
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.