Who engages? Machine learning insights into digital mindfulness-based intervention for generalized anxiety disorder.

Clinical Snapshot

Key Findings

• P Value: p < .001

• Effect Size: Cohen's d = 1.447 favouring MEMI over self-monitoring

• Primary Outcome: Two-week treatment engagement (log-transformed prompt completion)

• Nnt Or Sensitivity: Predictive model R² = 82.7% with RMSE = 0.780

• Confidence Interval: Not reported in abstract

Clinical Application

Digital delivery makes intervention accessible, but requires smartphone/device access and technical literacy Relevant to Australian digital mental health initiatives and telehealth expansion, though cost-effectiveness for PBS listing unclear without clinical outcome data Adults with GAD suitable for digital mental health interventions, particularly those with specific baseline characteristics identified by the predictive model

Abstract

BACKGROUND: Although mindfulness ecological momentary interventions (MEMI) appear effective in alleviating worry symptoms, treatment engagement remains suboptimal. Determining baseline variables of MEMI over self-monitoring placebo (SM) can inform tailored interventions for individuals with generalized anxiety disorder (GAD). METHOD: Machine meta-learning methods (ML) were applied to predict two-week engagement (log-transformed number of prompts completed) among individuals randomized to MEMI or SM (N = 110). Sixteen baseline variables comprised the predictor set: clinical, demographic, process, and executive functioning (EF) factors. Random forest using a five-fold nested cross-validation approach mitigated overfitting. X-learner meta-algorithms estimated conditional average treatment engagement (CATE). Shapley additive explanations evaluated relative importance. RESULTS: The 16-predictor model displayed strong predictive performance (R-squared [R2] = 82.7%; root mean squared error [RMSE] = 0.780; mean absolute error [MAE] = 0.512). The top-10 predictor model also yielded good predictive performance (R2 = 82.1%; RMSE = 0.547; MAE = 0.307). As predicted by the CATE analysiscate, participants had the highest treatment engagement when assigned to MEMI instead of SM (d = 1.447, p < .001). The following baseline variables predicted more engagement with MEMI over SM: lower GAD severity, inhibition response time (RT), and EF errors, higher attentional control, empathy, and verbal fluency (capitalization theory); lower mindfulness, and treatment expectancy, poorer working memory, and higher set-shifting RT (compensation model). LIMITATIONS: The small sample size, single engagement metric, and brief duration might constrain generalizability. DISCUSSION: Integrating robust ML approaches could optimally identify prescriptive predictors of engagement to brief digital mental health interventions to inform targeted treatments. TRIAL REGISTRATION: ClinicalTrials.gov ID (NCT04846777).

References

1. 1.Zainal, N. H., & Newman, M. G. (2026). Who engages? Machine learning insights into digital mindfulness-based intervention for generalized anxiety disorder. Journal of Affective Disorders. https://doi.org/10.1080/17437199.2025.2547610