AI Model Predicts Risks and Potential Causes of Adolescent Mental Illness
Contact
DURHAM, N.C. -- An artificial intelligence (AI) model developed by Duke Health researchers accurately predicted when adolescents were at high risk for future serious mental health issues before symptoms become severe.
Unlike prior models that primarily rely on existing symptoms, the AI model identified underlying causes, such as sleep disturbances and family conflict, that could be used to prescribe interventions. That capability could greatly expand access to mental health services, with assessments and care available through primary care providers.
“The U.S. is facing a youth mental health crisis -- nearly half of teens will experience a mental illness,” said Jonathan Posner, M.D., professor in the Department of Psychiatry and Behavioral Sciences at Duke and senior author of a study appearing March 5 in Nature Medicine.
“Despite this crisis, the U.S. has a critical shortage of mental health providers,” Posner said. “Our AI model could be used in primary care settings, enabling pediatricians and other providers to immediately know whether the child in front of them is at high risk and empowering them to intervene before symptoms escalate.”
Posner and colleagues -- including data scientists Elliot Hill and Matthew Engelhard, M.D., Ph.D., in Duke’s Department of Biostatistics & Bioinformatics -- analyzed psychosocial and neurobiological factors associated with mental illness using data from the ongoing ABCD study. The study has conducted psychosocial and brain development assessments of more than 11,000 children over five years.
Using artificial intelligence, the researchers built a neural network – an AI model that mimics brain connections -- to predict which children would transition from lower to higher psychiatric risk within a year. That model is then used to score a questionnaire that ranks responses from the patient or parent about current behaviors, feelings and symptoms, to predict the likelihood of an escalation.
The model was 84% accurate in identifying patients in the study who went on to have escalating illness within the next year.
Importantly, the Duke researchers analyzed an alternative model that identified the potential mechanisms that might lead to or trigger worsening mental illness. With an accuracy rate of 75%, the new modeling system’s ability to identify underlying causes gives it a unique ability to alert doctors and families to potential interventions.
“It’s a much easier task to say a child who has a fairly high symptom burden of mental illness is going to be ill in a year, than to determine that a child has all of these underlying risk factors for mental illness and is going to become ill,” Hill said. “It’s important to leverage that information to design an intervention for that child.”
Among the most common underlying causes of escalating illness are sleep disturbances, problematic behaviors, adverse events, family mental health history, and family conflict; of those, sleep disturbances emerged as the most powerful predictor of future psychiatric illness.
“It’s important to note that the model does not prove it's a child’s sleep disturbances that are causing an elevated risk, but it does suggest that's one of the limited number of malleable factors that seems to be associated with high rates,” Engelhard said.
The authors said the model demonstrates a way to reach a broader population of young patients with a tool that primary care practitioners could use to assess a child’s mental health risk using simple questionnaires.
“Primary care doctors often do not have the time to conduct a detailed psychiatric assessment, making it difficult to identify which children need early intervention. This AI model would automate the process, analyzing the data in real-time and providing the doctor with a simple output indicating the child’s risk level,” Posner said.
In addition to Posner, Hill and Engelhard, study authors include Pratik Kashyap, Elizabeth Raffanello, Yun Wang, Terrie E. Moffitt, and Avshalom Caspi.
The study received funding support from the National Institute of Mental Health (K01-MH127309); the National Institute on Aging (R01-AG032282 and R01-AG069939); the National Center for Advancing Translational Sciences (UL1 TR002553); the Medical Research Council (MR/P005918/1).
The content of this publication does not necessarily represent the official views of the NIH.