Machine learning looked for the teenagers school mindfulness helps. It barely found them.

When a universal prevention programme fails to beat its control condition, the standard rescue is personalisation: the programme worked, the argument goes, but only for some students, and the average washed them out. Find those students and the programme becomes useful again.

A secondary analysis of the MYRIAD trial, published in JAMA Psychiatry, tested that rescue directly on school-based mindfulness training — and the rescue did not arrive [s1].

The setup

MYRIAD was a cluster randomised trial run from October 2016 to July 2018 in 84 secondary schools, enrolling 8,376 adolescents who were 11 to 13 years old at baseline (mean age 12.2 years; 54.9% female, 43.2% male) [s1]. Students received either mindfulness training — core mindfulness skills taught through psychoeducation, class discussion and practice — or standard social-emotional learning as usual [s1]. The outcome was change in depressive symptoms on the Center for Epidemiologic Studies Depression scale [s1].

The authors note that evidence of effectiveness for school-based mindfulness training is mixed [s1]. The purpose of this analysis was to find out whether a data-driven algorithm, built from baseline characteristics alone, could identify which adolescents would benefit.

What the models did

Two approaches were trained using school-level nested cross-validation: a causal forest and an elastic net regression [s1]. Each produced a personalised advantage index — a per-student estimate of expected benefit from mindfulness training relative to teaching as usual [s1].

The causal forest showed acceptable calibration, with a mean best linear predictor slope of 0.78 (SE 0.15) [s1]. The elastic net's predictive performance was modest: r = 0.29, R² = 0.09, root mean square error 10.3 [s1].

Both models did identify a subset of adolescents predicted to benefit. The size of that benefit is the finding. Group differences in outcomes were negligible: Cohen's d of 0.07 (95% CI, 0.02 to 0.12; P = .007) for the causal forest, and 0.08 (0.02 to 0.13; P = .004) for the elastic net [s1].

Both are statistically significant. Neither is clinically meaningful. The authors say so plainly, describing the result as a subgroup with statistically detectable but clinically trivial differential intervention response [s1].

Why the p-values are misleading here

With 8,376 students, very small differences clear conventional significance thresholds easily. A d of 0.07 means the average student in the predicted-to-benefit group ended up a fraction of a standard deviation better off than a comparable student who was not. It is not a signal a school could act on, and it is not a basis for allocating one programme to one child and a different programme to another.

The features the models leaned on are worth noting. The causal forest's top predictors were symptom severity — low-to-moderate depression and anxiety predicted greater benefit — along with several school-level factors that behaved non-linearly [s1]. The elastic net emphasised school characteristics with minimal differentiation between students [s1].

That the strongest signal came from the school rather than the student is itself informative about where variation in these programmes lives.

The wider pattern

Personalisation is the same promise being made for digital youth mental health tools, and the implementation evidence there is similarly thin. A systematic review published in Internet Interventions searched for disseminated digital interventions for child and adolescent anxiety or depression that reported implementation outcome data, and found nine peer-reviewed articles covering seven different interventions [s2]. Use of implementation science was inconsistent, and use of implementation frameworks and models was minimal [s2]. The interventions were effective for and adopted by community users, but dropout rates were high [s2].

The common thread is not that these approaches do nothing. It is that the step between a trial result and a working programme in a real school keeps failing to be studied, and the assumption that better targeting will close the gap has now been tested once, at scale, and did not hold.

What to watch

The authors' conclusion is a caution about ambition rather than a verdict on mindfulness: achieving clinically useful personalisation in universal school-based prevention programmes faces substantial challenges [s1]. That applies beyond mindfulness. Any universal programme proposing to solve a null average result by finding responders now has to explain why its targeting will work where this did not.

This article is informational and is not medical advice.