From the Editor

The MYRIAD Trial was ambitious, involving more than 8 300 adolescents at 84 schools, with the aim of preventing depression and improving mental well-being by teaching mindfulness through a universal school program. The only catch? There was no difference in outcomes at one year.

Would it be possible to identify adolescents who would benefit from mindfulness? Christian A. Webb (of Harvard University) and his co-authors attempt to answer that question, using AI. And so, a longstanding objective, prevention, was joined with a modern method, machine learning. In the first selection, a paper from JAMA Psychiatry, the authors detail a secondary analysis using two complementary machine learning approaches and the MYRIAD Trial data. “This study found that analyses using machine learning identified a subgroup of participants with a 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.” We consider the paper and its implications.

In the second selection, from the Journal of the American Academy of Child & Adolescent Psychiatry, Alison Athey (of Johns Hopkins University) and her co-authors evaluate the impact of child access prevention laws on youth suicide deaths by firearms. They drew on more than 30 years of mortality data from the Centers for Disease Control and Prevention. “Laws that require families to store firearms unloaded and secured in a locking device appear to effectively prevent youth suicide deaths and firearm-related youth deaths by accident and homicide.”

And in this week’s third selection, Dr. Scott Monteith (of Michigan State University) and his co-authors write about generative AI and adolescents for The British Journal of Psychiatry. They note a surge in use – some 80% of British teens use generative AI – and consider problems, from cyberbullying to mental healthcare. “There is a need to increase awareness of how GenAI may have a negative impact on the mental health of teenagers.”

DG

Selection 1: “Predicting Adolescent Response to School-Based Mindfulness: A Secondary Analysis of the MYRIAD Trial”

Christian A. Webb, Boyu Ren, Verena Hinze, et al.

JAMA Psychiatry, 18 February 2026  Online First

Depression prevention research has grown considerably since early trials in the 1980s and 1990s, with momentum accelerating in recent years. Schools have emerged as a particularly promising venue for scalable delivery of depression prevention programs, as they provide access to nearly all children, across socioeconomic, racial, and ethnic backgrounds, offering a unique opportunity for broad-reaching interventions. A recent systematic review of school-based depression prevention programs concluded that, although these interventions may produce modest average benefits, the literature is marked by substantial heterogeneity, inconsistent findings, and concerns about methodological rigor and publication bias…

One approach that has garnered considerable attention within this broader effort is school-based mindfulness training (SBMT). SBMT teaches core mindfulness skills that are thought to buffer against the development of depression and related problems. To rigorously evaluate its effectiveness in real-world settings, the My Resilience in Adolescence (MYRIAD) trial was conducted: a cluster-randomized clinical trial… comparing SBMT with teaching standard social-emotional education (teaching as usual, TAU). Primary results showed no overall benefit of SBMT in reducing depression risk or enhancing social-emotional-behavioral functioning or well-being.

So begins a paper by Webb et al.

Here’s what they did:

  • They conducted a secondary analysis of the MYRIAD Trial.
  • That trial compared SBMT – which taught mindfulness skills through psychoeducation, class discussion, and practices – and standard social-emotional learning (teaching as usual).
  • They applied “two machine learning approaches, causal forests and regularized regression, to predict differential responses to SBMT vs TAU within the MYRIAD sample. Models were evaluated using nested (as opposed to regular, nonnested) 10-fold cross-validation to provide out-of-fold estimates of generalizability.” They used baseline student, teacher, and school-level characteristics, aiming “to identify which adolescents are most likely to benefit from SBMT, which may fare better with TAU, and which may experience little or even negative impact from either approach. By uncovering these individual differences in treatment response, we hope to advance more personalized, data-driven guidance that could translate into more targeted recommendations for school-based mindfulness programs.”
  • Main outcome: “Change in depressive symptoms from preintervention to postintervention measured by the Center for Epidemiologic Studies Depression scale.”

Here’s what they found:

  • There were 8 376 adolescents.
  • Demographics. The mean age was 12.2 years; most participants were female (54.9%).
  • Causal forest (CF) and elastic net regression (ENR) models. “CF showed acceptable calibration (mean [SE] best linear predictor slope = 0.78 [0.15]), while ENR demonstrated modest predictive performance (r = 0.29; R2 = 0.09; root mean square error = 10.3).”
  • Benefit. “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…; ENR: d = 0.08…).”
  • Prediction. “Top predictive features from the CF model were symptom severity (eg, low-to-moderate depression and anxiety predicted greater SBMT benefit) and several school factors with nonlinear patterns. ENR emphasized school-level characteristics with minimal differentiation.”

A few thoughts:

1. This is an interesting study published in a major journal.

2. The key finding in five words: machine learning didn’t really work. To add more detail, they were able to identify a subgroup but it was “statistically detectable but clinically trivial.”

3. For a more in-depth discussion of this paper, Dr. John Torous’ podcast is excellent. It includes a clear explanation of the difference between CF and ENR machine language approaches (spoiler alert: like the difference between a bronze and a gold at the Olympics). You can find it here: https://edhub.ama-assn.org/jn-learning/audio-player/19039530.

4. The authors justify the tact: “[the MYRIAD Trial] found no average benefit of universal SBMT over TAU in reducing adolescent depressive symptoms. Rather than suggesting prevention efforts should be abandoned, these findings highlight the need for more targeted or personalized strategies, particularly given the complex, multifactorial nature of adolescent depression risk. One explanation for these findings is that ‘1-size-fits-all’ universal interventions may obscure meaningful individual differences in response, with null average effects masking benefits (or harms) for specific subgroups.”

5. Will evolving AI eventually allow us to find the needle in the haystack with such datasets? Is the problem with the MYRIAD Trial itself (involving too many youth and trained-up teachers, rather than experts) making analysis futile to both humans and machines?

The full JAMA Psychiatry paper can be found here:

https://jamanetwork.com/journals/jamapsychiatry/article-abstract/2844884

Selection 2: “A National Evaluation of the Impact of Child Access Prevention Laws on Rates of Youth Suicide and Other Youth Firearm Deaths”

Alison Athey, Paul S. Nestadt, Megan L. Rogers, et al.

Journal of American Academy of Child & Adolescent Psychiatry, August 2025

Suicide is a leading cause of youth firearm mortality in the United States. Nearly one-third of youth firearm injury deaths are the result of suicide, and firearms are used in about half of all youth suicides. Population-level shifts in firearm ownership impact youth suicide rates. Children increasingly have access to firearms in their homes, as rates of firearm ownership increased during the COVID-19 pandemic. In line with increased firearm purchasing among parents, rates of firearm suicide increased by 68% in youth ages 10 to 14 and by 45% in youth ages 15 to 24…

Firearm access is a risk factor for youth suicide. Youth with access to firearms report higher rates of suicidal ideation, planning, and attempts compared with youth who do not have access to firearms. Firearm access during childhood may increase risk for suicidal thoughts and behaviors across the life span; adults who report that there was a firearm in their childhood home report significantly higher rates of suicidal ideation and suicide attempts compared with adults who deny exposure to firearms in childhood. More research is needed; however, it is possible that firearm access increases the capability for suicide in youth by increasing their familiarity with the methods for killing or decreasing their fear of lethal injuries. Firearm access also increases the risk that suicidal youth die during a suicidal crisis…

So begins a paper by Athey et al.

Here’s what they did:

  • They conducted a national evaluation of Child Access Prevention laws by drawing on nationally representative mortality data from 1990 through 2020 from the Centers for Disease Control and Prevention. The data was “disaggregated by state and year.”
  • Youth were included ages 1 to 17 years.
  • They analyzed CAP laws which “fall into 2 major categories: negligent storage of firearms policies that regulate how firearms are stored in households with children and reckless provision of a firearm to a minor policies that impose liability on firearm owners who provide youth with firearms that are used to harm others.” They compared rates in states with different CAP law provisions, examining whether more stringent requirements correlated with changes in adolescent firearm suicide rates.
  • Main outcomes: firearm suicide, nonfirearm suicide, firearm homicide, and firearm unintentional injury death rates among youth ages 1 to 17.

Here’s what they found:

  • Suicides. States with Child-Access Prevention laws had lower rates of youth firearm suicide than states without them.
  • Laws. Not all laws were equal; the strongest associations were seen in states that required firearms to be stored unloaded and locked (medium-to-large effects), not merely laws that imposed penalties after a child gained access.
  • Firearm suicides. The reductions were specific to firearm suicides. The authors did not find evidence that youth simply shifted to other methods. (!)
  • Statistical analyses. The findings held across multiple statistical models and sensitivity analyses.

A few thoughts:

1. This is an important study drawing on a solid dataset and published in a major journal.

2. The main finding in three words: the laws worked. More: “Our analyses show that policies that govern the storage of firearms in homes with children, but not laws that govern the provision of firearms to children, are associated with significant small-to-moderate reductions in youth firearm suicide rates.”

3. The percentage reductions were modest – but applied nationally over three decades, they translate into a meaningful number of lives saved.

4. They write: “Our findings support the growing body of literature that demonstrates that people who are prevented from accessing their suicide method of choice do not substitute another suicide method.” Restriction of means appears to be an important part of a successful suicide prevention strategy, applicable to, say, firearms in the US and pesticides in Sri Lanka.

5. Like all studies, there are limitations. The authors note several, including: “the low base rate of certain types of youth mortality may obscure their association with CAP firearm storage and reckless provision laws.”

The full JAACAP paper can be found here:

https://www.jaacap.org/article/S0890-8567(24)01991-9/fulltext

Selection 3: “Increasing use of generative artificial intelligence by teenagers”

Scott Monteith, Tasha Glenn, John R. Geddes, et al.

The British Journal of Psychiatry, 26 January 2026  Online First

A few years ago, many products that used artificial intelligence were based on machine learning models that made predictions from large sets of example data. For example, machine learning models have been used to predict whether an X-ray shows signs of a tumour.Recently, generative artificial intelligence (GenAI) models were developed, which create new data similar to the input data rather than making predictions. GenAI models that learn to produce new text from input text are labelled large language models (LLMs)… LLMs can be used to write articles and reports, create chatbots, summarise documents, translate between languages and generate software code. Some GenAI LLM models generate images, video and audio… The use of GenAI by teenagers has grown rapidly, used by nearly 80% of British teenagers in 2023 and 70% of US teenagers in 2024. Teenagers are likely to use GenAI for homework help (53%) but also to fend off boredom (42%).

So begins a paper by Monteith et al.

Teenage use of GenAI

“Adolescents utilise GenAI to write essays and reports, or create videos for social sharing. Many teens are using GenAI without telling their parents or teachers. While 50% of children age 12–18 have used GenAI for school, only 26% of parents are aware of such use. However, the potential consequences of over-reliance on GenAI may have an impact on critical thinking and creativity. Many teens easily believe GenAI output and treat it as if they were conversing with another human due to the human-like tone, aura of confidence and pattern-matching giving the convincing appearance of understanding and responding to what was said…

“While some children age 10–12 identify cultural, gender and racial biases in responses from GenAI, children may not be sufficiently critical of GenAI actions and responses. Children may be unaware that GenAI can make basic errors, such as ChatGPT giving an incorrect list of the states in the USA. Children and teenagers may be completely unsuspecting that GenAI can create coherent but inaccurate comments, referred to as hallucinations. GenAI may create harmful information that perpetuates historically biased stereotypes.”

Malicious use of GenAI LLMs

“GenAl LLMs can be used to alter real images to create fake images and create videos to deceive. Fake videos together with GenAI LLM chatbots can generate audio from a text script in any language or voice. GenAI LLM technology allows very sophisticated fake products to be created, commonly called deepfakes. If GenAI LLM generated images, video, audio and text are targeted at specific individuals for the purpose of harassment, it constitutes cyberbullying.”

Use of GenAI apps for mental healthcare

“The use of GenAI apps for healthcare, or wellness apps, can be risky. The GenAI app may not be able to recognise signs of mental illness. When used for mental healthcare, the patients may not be aware that the app is not a real person and does not have the emotional foundation for a caring relationship and is not capable of providing professional therapy.Some young people prefer human responses rather than GenAI responses for sensitive topics such as relationships and suicidal thoughts.”

They close with some recommendations:

“Steps are needed to teach children about the limitations of GenAI and how to differentiate fact from fiction.” They also suggest further research.

A few thoughts:

1. This paper is concise, thoughtful, and very timely.

2. This stat is worth repeating: half of adolescents use AI for school, yet only one in four parents is aware.

3. The authors conclude by suggesting that we teach kids about the limitations of AI. That seems reasonable. Should we also be teaching their parents?

The full BJPsych paper can be found here:

https://www.cambridge.org/core/journals/the-british-journal-of-psychiatry/article/increasing-use-of-generative-artificial-intelligence-by-teenagers/1F540923A9792823E18E8A2E2FE15859

Reading of the Week. Every week I pick articles and papers from the world of Psychiatry.