Highlights
The MYRIAD trial secondary analysis demonstrates that machine learning approaches, while technically feasible, yield clinically trivial predictions for personalizing school-based mindfulness training in adolescents. Low-to-moderate baseline symptom severity emerged as the most consistent predictor of intervention benefit, yet even identified subgroups showed minimal outcome differences. These findings challenge the feasibility of personalized prevention approaches in universal school-based mental health programs.
Background: The Burden of Adolescent Depression
Depression most commonly emerges during adolescence, marking a critical window for preventive intervention. School-based mindfulness training (SBMT) has emerged as a promising scalable approach, offering the potential to reach large populations of young people before clinical symptoms fully manifest. Despite widespread implementation, evidence for SBMT effectiveness has remained inconsistent, prompting researchers to explore whether a more personalized approach might yield better outcomes. The search for predictive biomarkers and baseline characteristics that could identify individuals most likely to benefit from such interventions represents a paradigm shift toward precision medicine in mental health prevention.
Study Design and Methods
The My Resilience in Adolescence (MYRIAD) trial was conducted across broadly representative secondary schools in England, Scotland, Wales, and Northern Ireland from October 2016 to July 2018. This secondary analysis utilized school-level nested cross-validation to train and evaluate machine learning models predicting individualized benefit from SBMT. The study included 8,376 adolescents aged 11 to 13 years at baseline from 84 UK secondary schools, with 4,509 (54.9%) female participants and 3,547 (43.2%) male participants. Data analysis was performed from April 2023 to October 2025.
Participants were randomly assigned at the school level to either SBMT—teaching core mindfulness skills through psychoeducation, class discussion, and structured practices—or to standard social-emotional learning as teaching as usual. The primary outcome was change in depressive symptoms from preintervention to postintervention, measured using the Center for Epidemiologic Studies Depression scale (CES-D). Two machine learning approaches were employed: causal forest (CF) and elastic net regression (ENR), both computing personalized advantage index scores that quantified the individual expected benefit from SBMT compared to teaching as usual.
Key Findings
Model Performance and Calibration
The causal forest model demonstrated acceptable calibration with a best linear predictor slope of 0.78 (SE 0.15), indicating reasonable consistency between predicted and observed outcomes. However, the elastic net regression model showed modest predictive performance with a correlation of 0.29, an R² of 0.09, and a root mean square error of 10.3. These metrics suggest that while the models could detect some signal in the data, substantial variance in individual outcomes remained unexplained.
Intervention Response Prediction
Both the CF and ENR models identified subsets of adolescents predicted to benefit from SBMT. However, when these predictions were tested, the group differences in outcomes proved negligible. For the CF model, the effect size was d = 0.07 (95% CI, 0.02-0.12; P = .007), while the ENR model yielded a similar effect size of d = 0.08 (95% CI, 0.02-0.13; P = .004). Although these differences reached statistical significance given the large sample size, the clinical magnitude of these effects is essentially negligible in practical terms.
Top Predictive Features
The causal forest model identified symptom severity as the primary predictor of intervention benefit. Notably, low-to-moderate depression and anxiety at baseline predicted greater SBMT benefit, suggesting a potential sweet spot for intervention targeting. Several school-level factors also emerged as important predictors, though these showed complex nonlinear patterns that complicate straightforward interpretation. The elastic net regression model placed greater emphasis on school-level characteristics while providing minimal differentiation at the individual student level.
Expert Commentary: Implications and Limitations
These findings illuminate the substantial challenges inherent in achieving clinically useful personalization within universal school-based prevention programs. The modest predictive performance and minimal effect sizes suggest that the heterogeneity of treatment effects in this population may be too small to detect with current methodologies, or that the mechanisms underlying SBMT benefit operate differently than anticipated.
Several limitations warrant consideration. First, the study relied on self-reported symptom measures, which may introduce measurement noise that limits predictive accuracy. Second, the 84-school sample, while substantial, may not capture the full diversity of educational and cultural contexts where SBMT might be implemented. Third, the relatively short follow-up period may miss longer-term trajectories of benefit or harm that could inform more nuanced predictions.
Furthermore, the nonlinear patterns observed in school-level predictors suggest that contextual factors may influence intervention response in complex ways that linear models struggle to capture. This raises questions about whether current machine learning approaches are appropriately suited for mental health prevention contexts where outcomes are influenced by intricate interactions between individual characteristics and environmental factors.
From a clinical perspective, these results do not necessarily indict SBMT as an intervention—rather, they highlight that the promise of precision medicine in preventive psychiatry remains largely unrealized for universal programs. Targeted approaches focusing on higher-risk populations might yield more substantial effects and clearer prediction signals, though this would fundamentally alter the universal prevention model.
Conclusion and Future Directions
The MYRIAD trial secondary analysis represents the most rigorous examination to date of personalized prediction for school-based mindfulness interventions in adolescents. While machine learning successfully identified a subgroup with statistically detectable differential response, the clinical relevance of this achievement remains questionable given effect sizes that would have negligible real-world impact.
The finding that low-to-moderate symptom severity predicted greater benefit offers a potentially actionable insight: adolescents with subclinical but elevated symptoms may represent the optimal target population for SBMT. Those with no symptoms have limited room for improvement, while those with more severe symptoms may require more intensive, individualized interventions beyond what universal programs can provide.
Future research should explore alternative modeling approaches capable of capturing complex individual-environment interactions, consider longer-term follow-up to identify delayed treatment effects, and examine whether targeted rather than universal delivery models might achieve more meaningful personalization. The field may also benefit from integrating multimodal data sources—genetic, neuroimaging, ecological momentary assessment—to improve prediction accuracy beyond what baseline characteristics alone can provide.
Trial Registration: isrctn.org Identifier: ISRCTN86619085
References
Webb CA, Ren B, Hinze V, et al. Predicting Adolescent Response to School-Based Mindfulness: A Secondary Analysis of the MYRIAD Trial. JAMA Psychiatry. 2026;83(4):389-398. PMID: 41706471.
