Wearable sensors might soon be able to tell the difference between opioid craving and everyday stress - the two currently look nearly identical to a smartwatch.
Researchers built RETRACE, a machine-learning framework that reads wearable physiology, think heart rate and related signals, to flag opioid craving without mistaking it for garden-variety stress. The core problem is that stress produces loud, reliable signals on wearables, while craving is faint and buried inside that same stress response. The team's fix uses a person's psychological resilience as context: since resilience cannot be read directly from a short sensor window, RETRACE infers it from proxies like how fast someone's heart rate recovers after stress and how they narrate their own memories. A dual-encoder setup then splits general stress physiology from person-specific craving interpretation, using gating and fusion to personalize predictions without needing craving labels from the target user.
On a new multimodal dataset combining wearable data, stress and craving annotations, and personal narratives, RETRACE beat the strongest baseline by up to 7 percentage points in leave-one-subject-out tests - meaning it generalized to people it never trained on, the hard part for any wearable health tool. That distinction matters for opioid use disorder specifically, where a false alarm or missed craving could shape whether someone gets a timely check-in instead of nothing at all. Most consumer wearables already track heart rate variability; the gap has been turning that noisy signal into something clinically useful for one person, not just a population average.
It's still a research paper built on a single custom dataset, not a product - and resilience proxies like memory-recall exercises aren't something your smartwatch app is going to run before breakfast. Leave-one-subject-out testing is a reasonable stand-in for whether this works on a stranger, but it's not the same as deployment in the wild.