National Institute on Alcohol Abuse and Alcoholism (R0 1AA031762)
John Curtin (PI)
Direct Costs: $1,751,304
Status: Active. 08/20/2024 – 07/31/2029
Clinician-delivered relapse prevention interventions for alcohol use disorder (AUD) are effective, but well-known barriers prevent both their delivery to the vast majority who need help and also increase mental health disparities. Digital therapeutics can address these barriers by providing 24/7 access to interventions and other supports that are modestly efficacious generally and for marginalized or otherwise under-served groups. Unfortunately, the benefits from digital therapeutics may be constrained because engagement is often not sustained or matched to patients’ needs. Patients may not know when they should use a digital therapeutic, which intervention and other support modules are best for them, and more precisely, which are best for them at that moment given the dynamic nature of relapse and recovery. Successful digital therapeutic use, and recovery more generally, require careful monitoring of behaviors, interpersonal interactions, lifestyle, and other risk factors. Yet this monitoring is difficult given the complex, non-linear dynamics of relapse.
Machine learning prediction models powered by personal sensing can provide this monitoring. Curtin et al. (PI) has used such sensing in an XGBoost machine learning model to predict individual, temporally-precise probabilities (updated hourly) for future lapses back to drinking in the next 24 hours with high sensitivity and specificity. This model can also identify the most influential relapse-related processes at play based on important model features (predictors) for that person at that moment in time.
“Smart” digital therapeutics that are enhanced by this or other emerging machine learning prediction models may guide patients to sustain engagement with the specific interventions and supports that are most personally risk-relevant and therefore most effective. However, we must first determine how best to provide model feedback such that patients use this information and follow its recommendations. Some model transparency66 can improve perceptions of embedded machine learning models. Transparency may also promote patient learning and insight to yield additional clinical benefits directly. However, complex or otherwise unnecessary information can undermine trust in these models.
Our broad goal is to develop a machine learning guided engagement message system that can be added to any smart digital therapeutic to increase risk-relevant engagement and improve clinical outcomes. Following the Multiphase Optimization Strategy (MOST), we manipulate four candidate components of daily “engagement messages” that convey transparent, individualized, risk-relevant information from our machine learning lapse prediction model to participants. These message components include: 1) lapse probability, 2) lapse probability change, 3) important model features, and 4) a risk-relevant module recommendation. We pursue three aims:
Aim 1: Test the effects of message components on lapse risk-relevant engagement with a smart digital therapeutic. We use a MOST factorial experiment to determine which of the four message components encourage individuals to use information from the machine learning model to increase their engagement with intervention and support modules that are personally risk-relevant.
Aim 2: Test the effects of message components on two clinical outcomes (drinking days and heavy drinking days). In addition to the Aim 1 engagement outcome, we test component effects on clinical outcomes because the risk-relevant information from our machine learning model may provide direct benefits to participants through mechanisms other than engagement (e.g., information about relapse processes highlighted by important model features may promote adaptive lifestyle and behavioral adjustments that do not require digital therapeutic use). Message components that either increase risk-relevant engagement (Aim 1) or improve clinical outcomes (Aim 2) will be recommended for use in future smart digital therapeutics.
Aim 3: Precisely characterize the relationship between digital therapeutic engagement and clinical outcomes. We use daily measures of digital therapeutic engagement and clinical outcomes to test a) if increased overall digital therapeutic engagement is associated with improved clinical outcomes, b) if the effect of engagement on clinical outcomes is stronger for risk-relevant vs. other digital therapeutic module use, and c) if the effect of engagement on clinical outcomes changes based on daily lapse probability (e.g., is engagement more important on days when lapse probability is high). This can establish engagement generally and risk-relevant engagement more specifically as important proximal mechanisms for the clinical benefits from smart digital therapeutic use. Exploration of the moderating role of lapse probability can guide subsequent research on the optimal frequency and timing of engagement messages.