AI predicts adolescent mental health risk before symptoms emerge

A new study published in Nature Medicine demonstrates that artificial intelligence can identify adolescents at high risk for serious mental health problems before symptoms become severe. This innovative model goes beyond simply looking at current symptoms; it identifies underlying factors, such as disruptions in sleep patterns and conflicts within families, that contribute to these risks. This capability opens up the possibility of significantly improving access to mental health support, potentially making assessments and early interventions available through primary care doctors.

Rates of mental illness among young people have increased considerably, placing even greater pressure on already stretched mental health services. A major obstacle in improving mental health care is the difficulty in pinpointing which young people are most vulnerable and at the highest risk of developing psychiatric conditions. Being able to accurately predict which individuals in the general population will develop mental health problems would allow for a more efficient distribution of resources aimed at prevention.

“The United States is facing a youth mental health crisis. Almost 50% of teens will experience some form of mental illness, and of those, two-thirds will not get support from a mental
health professional,” explained study author Elliot Hill (@elliotdhill), an AI Health Fellow at Duke University School of Medicine.

“We wanted to test if AI could be used to help detect which children are most at risk of worsening mental health. If we can predict who is at risk, we can better allocate mental health resources to patients that need it the most to reduce the demand on over-burdened providers.”

The scientists used data from a large, ongoing study called the Adolescent Brain and Cognitive Development Study, which includes over 11,000 children across the United States. This study collects information about various aspects of these children’s lives, including their social environments, behaviors, and brain development, over several years. The researchers used this extensive data to train computer models known as neural networks. These models are designed to learn complex patterns from large amounts of data. The aim was to see if these models could predict a teenager’s future mental health risk based on information collected earlier.

The research team created two main types of prediction models. One type, called a symptom-driven model, was trained to predict future mental health risk based on the symptoms teenagers were already showing. This approach is similar to how risk is often assessed currently.

The other type, called a mechanism-driven model, was designed to predict risk based on potential underlying causes of mental health issues, such as problems with sleep, family difficulties, and stressful childhood experiences. This model did not rely on current symptoms. Both models used questionnaires completed by the teenagers and their parents. Some models also incorporated brain scans, obtained through a process called magnetic resonance imaging, to see if brain measurements could improve predictions.

To measure mental health risk, the researchers used a concept called the “p-factor.” The p-factor is a way of measuring general mental health difficulties across different types of problems, such as anxiety, depression, and behavioral issues. Instead of focusing on specific diagnoses, the p-factor provides a single score that reflects an individual’s overall level of psychological distress. The research team divided the teenagers into four groups based on their p-factor scores, ranging from no risk to high risk. The computer models were then trained to predict which risk group a teenager would fall into one year later.

The artificial intelligence model was able to predict which adolescents would develop serious mental health issues with high accuracy. The model trained on existing psychiatric symptoms achieved an accuracy score of 0.84, while the model trained solely on underlying causes reached a score of 0.75.

The findings indicate that “AI models trained on psychosocial and behavioral questionnaires can accurately predict future mental health risk while simultaneously suggesting potential targets for intervention,” Hill told PsyPost. “Our model highlighted the importance of sleep quality and prosocial behaviors for predicting future mental health risk.”

Among the various factors analyzed, sleep disturbances emerged as the strongest predictor of future psychiatric illness. The impact of sleep problems on mental health risk was greater than that of adverse childhood experiences or family mental health history. Adolescents with significant sleep disturbances were far more likely to transition into the highest-risk group within a year. Other influential factors included family conflict and low levels of parental monitoring.

“In the literature, adverse childhood experiences and family mental health history are often thought to be dominant predictors of future mental health,” Hill said. “While these factors were still strong predictors in our model, the influence of sleep quality on mental health predictions was even stronger. This is a hopeful finding because this factor is modifiable through evidence-based behavioral interventions.”

Interestingly, the inclusion of brain imaging data did not improve the model’s performance. This suggests that simple psychosocial questionnaires—rather than expensive and difficult-to-access neuroimaging measures—may be sufficient for identifying mental health risk. The findings indicate that artificial intelligence models could be used in routine healthcare settings, such as pediatric clinics or schools, to flag at-risk adolescents before they develop severe psychiatric conditions.

The researchers acknowledged some limitations to their study. The data came from a general population of teenagers, not specifically from young people already seeking mental health treatment. Therefore, it will be important to test these models in clinical settings to ensure they work effectively for those seeking help. Future research should also explore ways to make these prediction tools even more practical and accessible. This could involve identifying the smallest set of questionnaire questions needed to maintain accuracy, reducing the burden on individuals taking these assessments.

“Though the ABCD study was a general sample of the US population, it is possible that clinical populations are systematically different from the general population,” Hill explained. “Thus, it is vital to test our model in clinical settings before deploying it at large. Therefore, we are working on a grant to test this model in a clinical setting. We are targeting urban areas in North Carolina, as there is a critical shortage of mental health care providers in these areas.”

“This project was a diverse multidisciplinary collaboration between machine learning researchers, psychologists, psychiatrists, and neuroscientists. It would not have been possible without the help of my amazing co-authors.”

The study, “Prediction of mental health risk in adolescents,” was authored by Elliot D. Hill, Pratik Kashyap, Elizabeth Raffanello, Yun Wang, Terrie E. Moffitt, Avshalom Caspi, Matthew Engelhard and Jonathan Posner.