Duke researchers receive $15M federal grant to expand AI model designed to predict mental illness
DURHAM, N.C. (AP) — A team at Duke University has secured a $15 million federal grant to expand an artificial intelligence model designed to predict mental illness in adolescents.
The Duke Predictive Model of Adolescent Mental Health (Duke-PMA), co-developed by Professor of Psychiatry Jonathan Posner, Assistant Professor of Biostatistics & Bioinformatics Matthew Engelhard and AI Health Fellow Elliot Hill, is an AI-based tool that assesses factors related to adolescent mental health.
The model is used to predict who is most likely to develop a mental illness within a year. It also identifies the key factors driving those predictions, offering the potential to guide targeted preventive interventions.
“In the way that psychiatry is currently practiced, it tends to be reactive, meaning we wait until someone’s developed a psychiatric illness, and then we institute treatment,” Posner said. “So (the model) would really be a paradigm change in psychiatry from a reactive to a proactive approach.”
The model achieved 84% accuracy in identifying adolescents of age 10 to 15 who are at risk for future serious mental health issues and maintained consistent performance across socioeconomic status, race and sex. This accuracy was achieved using only questionnaires, instead of expensive imaging or blood tests, making the model a highly scalable and accessible assessment tool.
The model maintained high accuracy when limited to factors that can be directly influenced through clinician intervention, such as sleep disturbances and family conflict. Its results could offer clinicians actionable insights to guide prevention and intervention strategies before illness develops.
“So a patient comes into their clinic, they do this quick assessment, and then the primary care doctor gets a report saying, this child in front of me has a 90% chance of developing an illness within a year, and these are the factors that are driving that prediction,” Posner said.
Securing the $15 million federal grant marks a turning point in the project’s development. “This is exactly the pathway to get it in (the clinicians’) hands and actually identify people early and connect them with services and support that can hopefully bend that trajectory,” Engelhard said.
The next phase of the project will enroll 2,000 adolescents from rural clinics in North Carolina, Minnesota and North Dakota.
“We wanted to go to places where the resources for mental health care are pretty limited across the board,” Posner said. “Having an automated tool like this, while it would be helpful virtually anywhere, would be particularly helpful in a rural setting, which doesn’t have the mental health resources that you’d see in an urban clinic.”
The team will conduct an observational study, using the Duke-PMA to assess participants and generate predictions. Families will be recontacted a year later for detailed psychiatric evaluations to determine whether the model’s predictions prove accurate.
The use of artificial intelligence in medicine may spark both excitement and unease, particularly when applied to sensitive areas like adolescent mental health. For one, to address the risk of false positives, Hill emphasizes that Duke-PMA is designed as a supportive tool, not a replacement for clinical judgment.
“We’re very serious about protecting patients’ privacy, both in the context of the study that we’re doing, as well as more broadly, going forward,” Engelhard said. “And so this is information that would be between you and your care providers.”
This approach attempts to balance innovation with caution, enhancing care while preserving essential human presence during clinical judgment.
“This type of research would not be possible unless you had people from lots of different disciplines collaborating together … I think Duke is unusually well positioned for that type of work,” Posner said.
