Wearable Physiology Predicts Opioid Craving with Resilience Guidance.

Detecting opioid craving from wearable physiological data is a significant challenge, particularly because craving signals are often subtle and entangled with more dominant stress responses. This research introduces RETRACE, a novel framework designed to improve the accuracy of subject-independent craving estimation in individuals with Opioid Use Disorder (OUD). RETRACE addresses the heterogeneity of craving by reframing detection as a "trait-conditioned physiological interpretation," meaning it uses individual resilience levels to guide the interpretation of physiological patterns. The framework employs a dual-encoder design: a frozen stress-pretrained encoder handles general physiological responses, while a resilience-conditioned craving encoder focuses on subject-specific craving interpretations. This combination, along with feature-level gating and representation-level fusion, allows for lightweight personalization without requiring target-user craving labels or extensive retraining. Evaluated on a new multimodal OUD dataset, RETRACE demonstrated up to a 7% absolute improvement over strong baselines, highlighting its potential for more effective and proactive interventions.