Predicting Adolescent Response to School-Based Mindfulness

Importance: Depression most commonly first emerges during adolescence, making early prevention critical. While school-based mindfulness training (SBMT) offers a scalable prevention approach with broad reach, evidence of its effectiveness is mixed, and there is a compelling case for a more personalized approach to prevention. Objective: To develop a data-driven algorithm from baseline characteristics to predict which adolescents are most likely to benefit from SBMT. Design: Secondary analysis of the My Resilience in Adolescence (MYRIAD) cluster randomized controlled trial conducted between 2016-2018. School-level nested cross-validation was used to train and evaluate machine learning models for predicting individualized benefit from SBMT. Setting: Eighty-four broadly representative UK secondary schools across England, Scotland, Wales, and Northern Ireland. Participants: A total of 8,376 adolescents aged 11-13 years at baseline were included. Students had to be enrolled in participating schools with parental consent and student assent. The sample was 55% female, 76% White ethnicity, with mean age 12.2 years. Interventions: SBMT teaching core mindfulness skills through psychoeducation, class discussion, and practices, compared to standard social-emotional learning (teaching-as-usual; TAU). Main Outcomes and Measures: Change in depressive symptoms from pre- to post-intervention measured by the Center for Epidemiologic Studies Depression Scale. Causal forest (CF) and elastic net regression (ENR) models computed Personalized Advantage Index scores quantifying individual expected benefit from SBMT vs. TAU. Results: CF showed acceptable calibration (mean Best Linear Predictor slope = 0.78, SE = 0.15), while ENR demonstrated modest predictive performance (r = 0.29; R² = 0.09; RMSE = 10.3). Both the CF and ENR models identified a subset of adolescents predicted to benefit from SBMT, but group differences in outcomes were negligible (CF: d = 0.07, p = .007; ENR: d = 0.08, p = .004). Top predictive features from the CF model were symptom severity (e.g., low-to-moderate depression and anxiety predicted greater SBMT benefit) and several school factors with non-linear patterns. ENR emphasized school-level characteristics with minimal differentiation. Conclusions and Relevance: Machine learning identified a subgroup with statistically detectable but clinically trivial differential intervention response. These findings highlight the substantial challenges in achieving clinically useful personalization in universal school-based prevention programs. Trial Registration: ISRCTN86619085