1 - (SYM 134) Identifying Predictors of Engagement to a Brief Mindfulness Ecological Momentary Intervention for Generalized Anxiety Disorder Using Explainable Artificial Intelligence (AI)

Background: Since engagement tends to be low toward digital mental health interventions such as mindfulness ecological momentary interventions (MEMIs), identifying predictors of engagement might inform targeted treatments. The present study thus harnessed explainable artificial intelligence (AI) methods to identify predictors of engagement with a 14-day MEMI app within a controlled trial.

Method: Participants with generalized anxiety disorder (GAD; N = 110) were randomized to a MEMI or self-monitoring app (SM). They completed baseline self-reports of treatment credibility, expectancy, trait attentional control, empathy, mindfulness, repetitive thinking, symptom severity of depression and GAD, socio-demographics, and performance-based neurocognition. Nested ten-fold cross-validated machine learning (ML) models were conducted with sixteen baseline predictors for both groups to predict engagement (defined as completing ≥ 80% of app prompts). Shapley additive explanations (SHAP) bee swarm plots enabled interpretation of the association between each predictor and outcome, adjusting for other predictors in the multivariate model.

Results: The final models for the top five predictors of engagement performed well in both the MEMI (area under the receiver operating characteristic curve [AUC] = .741) and the SM sample (AUC = .946). In the MEMI, one strength (higher treatment credibility), three weaknesses (lower working memory, trait empathy, and higher GAD severity), and being White were observed to predict better engagement probability. In the SM, three strengths (higher trait empathy, lower trait repetitive thinking, executive dysfunction), one weakness (lower trait attentional control), and being White were found to predict stronger engagement probability.

Discussion: By collecting diverse neurocognitive and self-reported data at intake, explainable AI can be harnessed with good multivariate predictive accuracy to identify which client with GAD would more likely engage with the MEMI or SM to inform treatment matching. Whereas clients with more deficits might more likely use and benefit from the MEMI over SM, aligning with the compensation model, the opposite occurred for SM, concurring with the capitalization model. Unique cultural adaptations and user-centered designs should be applied to potentially enhance engagement with MEMI or SM for GAD. If replicated, this approach might innovatively pave the way to improve psychotherapy delivery within stratified settings and reduce mental health treatment disparities.