Smartwatches and AI Predict Opioid Misuse Risk, UC San Diego Researchers Find

Researchers at the University of California San Diego (UC San Diego) have developed an innovative system utilizing smartwatches and machine learning to predict the risk of opioid misuse among individuals living with chronic pain. This system aims to offer continuous monitoring and enable timely interventions, addressing the critical limitations of periodic clinical assessments in capturing real-time fluctuations in risk.

Background

Opioid-related overdose deaths in the U.S. reached nearly 80,000 in 2023, with opioids contributing to the majority of drug-related fatalities globally. Individuals managing chronic pain with long-term opioid prescriptions face an increased risk of misuse. This risk is often linked to elevated stress, pain flare-ups, and cravings. Current clinical assessments typically occur periodically, which can miss crucial real-time changes in a patient's risk profile.

The system aims to provide continuous monitoring and enable timely interventions, addressing the limitations of periodic clinical assessments in capturing real-time risk fluctuations.

Research Team and Methodology

The study was led by Professor Tauhidur Rahman and Ph.D. student Yunfei Luo from the Halıcıoğlu Data Science Institute, alongside Professor Eric Garland from the UC San Diego School of Medicine.

The research involved developing a system that utilizes a commercially available Garmin Vivosmart 4 smartwatch to collect inter-beat interval data, which measures the subtle timing differences between heartbeats. This data is then used to estimate heart rate variability (HRV), a key indicator of how the nervous system responds to stress. Machine learning algorithms are subsequently applied to identify patterns associated with an increased risk of opioid misuse.

The study included 51 adults who had chronic pain and were undergoing long-term opioid therapy. Over an eight-week period, 10,140 hours of wearable data were collected from participants in their daily environments. Participants' risk levels were categorized using the Current Opioid Misuse Measure (COMM) questionnaire.

Personalized Risk Prediction

The system maps heart rate variability to opioid misuse risk through a two-step process:

1. Personalized Prediction of Stress, Pain, and Craving: Recognizing the individual nature of HRV patterns, the team trained personalized models instead of a universal predictor. A learning-to-branch technique was employed to cluster participants with similar characteristics, facilitating data-efficient and individualized predictions of stress, pain, and craving levels.

2. Estimating Misuse Risk from Daily Patterns: Researchers analyzed the evolution of stress, craving, or pain states over time using nonlinear dynamical analysis. The study observed that individuals at a higher risk for opioid misuse exhibited more repetitive daily patterns and tended to remain in states of high stress, pain, or craving, indicating lower entropy or reduced flexibility. In contrast, those taking opioids as prescribed showed greater fluctuation and adaptability, reflected as higher entropy.

Integration of Clinical Data

To further enhance prediction accuracy, the system incorporates existing medical record information. This includes demographics, prescription history, symptoms, and related conditions. Clinically trained language models convert textual data from these records into numerical summaries for the prediction model. The combination of smartwatch signals with this clinical context significantly improved the system's overall prediction performance.

Implications and Future Directions

This type of continuous monitoring is designed to support "just-in-time interventions," which could deliver assistance precisely when needed. The long-term objective of the research is to transition from periodic patient assessments to continuous, patient-friendly monitoring, aiming to enable earlier intervention against opioid misuse.

This innovative approach is intended to allow clinicians to detect risk shifts between clinic visits, facilitate timely interventions, reduce the need for constant self-reporting, and better target prevention strategies for chronic pain patients.

The study's findings were published in Nature Mental Health, and a U.S. utility patent application (US2025/016369) has been filed for the technology.