Contextualized daily prediction of lapse risk in opioid use disorder by digital phenotyping
National Institute on Drug Abuse (R01 DA047315)
John Curtin (Co-PI), Dhavan Shah (Co-PI), Xiaojin Zhu (Co-I), David Gustafson (Co-I), Randy Brown (Co-I), William Sethares (Co-I), Qunying Huang (Co-I)
Direct Costs: $2,238,540
Status: Funded. 08/01/2019 – 06/30/2024
Opioid use disorder is increasingly widespread, leading to devastating consequences and costs for patients and their families, friends, and communities. Overdose deaths are soaring, having more than quadrupled since 1999. Available treatments for opioid and other substance use disorders (SUD), while effective for initiating abstinence, are not successful at sustaining abstinence. The vast majority of people with SUD relapse within a year, and often much sooner. Critically, they often fail to detect dynamic, day-by-day changes in their risk for relapse and therefore do not adequately employ skills they developed or take advantage of support available through continuing care.
Well-established theoretical models indicate that SUD lapse risk is a dynamic, non-linear function of both distal, relatively static, patient characteristics and time-varying changes in proximal, precipitating risk factors. However, comprehensive, precise assessment of dynamic risk signals with high temporal precision has not been possible until very recently. Furthermore, innovative methods from predictive analytics have only begun to be applied to the lapse risk prediction problem and often rely on only a few self-reported and isolated set of risk signals rather than a diverse, unobtrusive, and dynamic set of signals situated within a rich environmental and social context.
The broad goals of this project are to train, validate, test, and deliver highly contextualized, lapse risk prediction models for forecasting day-by-day probability of opioid and other drug use lapse among people pursuing drug abstinence. These lapse risk prediction models will be integrated into the Comprehensive Health Enhancement Support System for Addiction (A-CHESS) mobile app, which has been established by RCT as a state-of-the-art mHealth system for providing continuing care services for substance use disorders. Such a risk prediction model can be used to alert people in recovery to apply available skills and seek services when they are most needed to prevent relapse. Day-by-day lapse risk probabilities could also be shared with A-CHESS connected continuing care providers, allowing them to offer efficient and effective stepped care across their case load.
To pursue these goals, we will recruit 480 participants with opioid use disorder using local and targeted national advertising in Facebook. We will enroll participants with 1 week to 6 months of opioid abstinence and follow them for 12 months. This will allow us to identify lapse risk signals that span early vs. later periods of recovery (from 1 week to 1.5 years post-cessation of use). We will recruit diverse participants (e.g., differing on race, gender, comorbid SUD and other psychopathology, SUD treatment use) across urban, suburban and rural settings to support model generalization.
We will measure well-established distal, static lapse risk signals (e.g., addiction severity, comorbid medical and psychiatric illnesses, chronic pain) on intake. We will also measure a range of proximal, time-varying risk signals through (1) continuous information gathering via smartphone sensors (e.g., GPS, microphone, camera) and (2) natural language processing of text messages, social media posts, and daily video check-ins. These proximal signals include subjective experience (e.g., ecological momentary assessments of affect, craving, stressors, risky situations), social network patterns and communications (e.g., voice call and text messaging, social media activity and content), acoustic and visual signal processing (facial and vocal displays from “video check-ins”) and health relevant behaviors (e.g., sleep quality, activity level, Medication-assisted and other treatment utilization) over 12 months for each participant. These dynamic risk signals will be modeled in a densely-sampled longitudinal intra-personal context. The predictive power of these signals will be further increased by anchoring them in a rich inter-personal context (e.g., contextualized voice call logs indicate frequency of calls to supportive vs. opioid using friends/family; likewise, GPS signal places the participant in the home of a Narcotics Anonymous sponsor or meeting vs. park associated with drug purchases). This contextualized mHealth approach provides the foundation for these Specific Aims:
1. Use machine learning methods to train, validate, and test an opioid lapse risk prediction model based on contextualized static and dynamic risk signals. This opioid lapse risk prediction model will provide participant specific, day-by-day probabilistic forecast of a lapse to opioid use among opioid abstinent individuals.
2. Use machine learning methods to train, validate, and test alcohol or other drug (AOD) lapse risk prediction models based on contextualized static and dynamic risk signals. Although we recruit participants with an opioid use disorder diagnosis, the large majority of these participants are expected to be poly-substance users pursuing abstinence from other drugs. We will leverage this opportunity to simultaneously develop separate lapse risk models for 1) alcohol, 2) stimulants, and 3) all drugs combined. This will allow us to identify signals that operate selectively for specific drugs vs. signals that predict generic risk for lapse for any drug.
As part of AIMS 1-2, we will compare prediction accuracy across models using all risk signals vs. simpler models that ignore specific signals or reduce their sampling frequency. These analyses will identify the minimal set of signals necessary for precise prediction to balance participant burden, implementation complexity and clinical benefits.
3. Add the lapse risk prediction models to the A-CHESS mobile app for real-time implementation in clinical care. Quantification of lapse risk probability is useful only if it is possible to use this information to support continued abstinence. The A-CHESS app provides an ideal platform within which to provide lapse risk probabilities to both the user and their A-CHESS connected clinicians to intervene “just in time.” Critically, A-CHESS already has real-time access to most of the data sources for our risk signals (e.g., GPS, surveys, ecological momentary assessments, discussion board posts, peer-to-peer messages) increasing the feasibility of this aim. Model parameters will also be documented in published papers to allow integration of risk prediction into other continuing care systems.