Digital Markers in Relapse and Recovery

National Institute On Drug Abuse

Investigators

Linked publications, trials & patents

Abstract

The study of relapse in substance use disorders (SUDs) continues to be marked by methodological variation, including differences in how relapse is defined, assessed, and modeled. Traditional approaches to relapse data collection include: (1) retrospective self-reports, where participants recall past relapse events and preceding factors; (2) prospective designs, where potential predictors are collected at baseline or at intervals and later analyzed in relation to relapse outcomes; and (3) near real-time reports, where participants provide information electronically close to the time of the relapse event. Direct behavioral observation in daily life has been limited due to challenges in obtaining consistent, high-quality data. Near real-time reports are particularly valuable because relapse-related factors such as self-efficacy, craving, stress, mood, and social interactions can shift rapidly, sometimes within minutes. This project addresses these challenges by using real-time and passive digital data to detect and predict relapse among individuals in treatment and recovery. During this fiscal year, we advanced our integration of social media language, smartphone sensor data, and wearable device data to generate predictive models of relapse and long-term recovery. Natural language processing and machine learning methods were applied to large-scale datasets, enabling the creation of models that forecast relapse risk and recovery trajectories. Passive data collection provided continuous behavioral measures, complementing and in some cases exceeding the resolution of traditional assessments. Notable accomplishments include publication of research demonstrating that language from social media platforms, combined with psychological self-reports, can predict U.S. county-level opioid poisoning mortality with greater accuracy than regional indicators or healthcare access measures.. We also validated the generalizability of predictive language models across different communication platforms and applied smartphone-based sensing in a clinical population with alcohol-associated liver disease and alcohol use disorder to estimate alcohol craving. Looking forward, these efforts are converging toward the development of an automated, continuous monitoring system that synthesizes inputs from multiple digital sources to produce daily relapse vulnerability scores. This system will be designed to operate as part of real-world, reproducible, and scalable intervention frameworks, directly supporting the development and deployment of evidence-based tools for healthcare providers and individuals engaged in treatment or recovery.