By combining information collected during healthcare encounters with data from the environment, the ‘trajectories’ team is using machine learning to identify children at highest risk of developing a mental illness over time.Credit: Cincinnati Children’s Hospital Medical Center
Young people are in the throes of a mental health crisis. They suffer from increasing rates of depression, anxiety and suicidal thoughts.
It is well established that adult-type mental disorders typically emerge early in life1,2, and there is evidence that the course of mental illness can be modified through early intervention3. But recognizing signs of impending psychiatric illness in young people is not always easy. A new decision-support tool being developed by Cincinnati Children’s Hospital Medical Center and its partners could help.
Detecting abnormal trajectories
The Mental Health Trajectories programme — the latest effort of the Decoding Mental Health Center at Cincinnati Children’s — is developing tools that combine information routinely collected by doctors with geographic, environmental and other data. Artificial intelligence (AI) then, over time, identifies kids at greatest risk of developing a mental illness for early care and intervention. Just as paediatric growth charts plot the course of children’s height and weight over time, these tools aim to do the same for the far more complex trajectories of anxiety, depression and suicide ideation.
Unpublished preliminary testing suggests early identification of anxiety is possible with this approach. Given the potential to transform the management of mental health, researchers are keen to validate these encouraging data in clinical settings.
“With successful validation, this trajectory approach will change the way we diagnose and treat mental health,” says John Pestian, a neuropsychiatric AI scientist who co-directs the Decoding Mental Health Center at Cincinnati Children’s. “If we can identify problems early and develop new approaches to care, we can reduce their impact now and during adulthood.”
Harnessing AI insights
The project builds on the work of Pestian’s lab over the past 20 years to develop machine-learning methods for identifying individuals at risk of suicide. Using speech samples from therapy sessions and notes from people who died by suicide, Pestian and colleagues showed that AI algorithms could detect suicidality with up to 90% accuracy4.
Pestian’s lab has also collaborated with paediatric neurologist and epileptologist Tracy Glauser, associate director of the Cincinnati Children's Research Foundation and co-director of the Decoding Mental Health Center, to develop an AI-based process for determining whether children with epilepsy are candidates for neurosurgery.
Epilepsy surgery is a potentially curative procedure typically reserved for children for whom anti-seizure medications are ineffective. Too often, Glauser says, physicians delay offering surgery to their patients. The longer someone with drug-resistant epilepsy waits for surgery evaluation, the longer they continue to experience seizures.
Glauser and Pestian developed a computer model that uses changes in neurologists’ notes about a patient over time. This time-series natural language processing enabled the algorithms to compare physicians’ written language between patients with well-controlled seizures and those with drug-resistant seizures who underwent epilepsy surgery. The model has been shown to identify surgery candidates years earlier than standard methods5. Validation, testing and biases analysis has been funded by the National Institutes of Health. The tool is now in use at Cincinnati Children’s.
Supercomputing power
Encouraged by these successes, Glauser and Pestian decided to apply similar techniques to some of the most pernicious problems in childhood mental health — anxiety, depression and suicidality. Cincinnati Children’s invested US$10 million to support this effort.
“We proved you can use validated algorithms with good data to start affecting care,” says Glauser. The goal now is to move beyond specialized clinics and develop “a mechanism by which paediatricians, community agencies and schools can have an early warning signal for mental health issues”.
Achieving that goal is no small task. To be accurate, the algorithm needs to be trained on huge data sets. The project is making use of records from 1.3 million patients cared for at Cincinnati Children’s, whose nine million visits generated both structured and unstructured data. Unstructured data requires natural language processing, a particularly complex technique. A standard computing cluster, Glauser recalls, would not be up to the job. “We realized that the size of the problem would require world-class supercomputing.”
The Summit supercomputer at the US Department of Energy's Oak Ridge National Laboratory has a peak performance of 200,000 trillion calculations per second, allowing for analyses that might otherwise take years to be performed in less than a day.Credit: Carlos Jones/Oak Ridge National Laboratory
They partnered with Oak Ridge National Laboratory (ORNL), the US Department of Energy’s largest science and engineering lab and home to the world’s most powerful computers. The laboratory’s Summit supercomputer can crunch through the Cincinnati Children’s data at a speed of up to 200,000 trillion calculations per second, allowing for analyses that might otherwise take years to be performed in less than a day. Pestian says this computational power “will allow us to provide clinicians with the information they need, when they need it”.
Gathering specialists from all corners
Glauser and Pestian aren’t alone in their quest. Their 'trajectories’ programme team includes more than two dozen scientists from Cincinnati Children’s as well as collaborators from ORNL, the University of Cincinnati and the University of Colorado. The partners’ expertise ranges from computer science to psychiatry, paediatrics and childhood development.
Working with the University of Cincinnati’s College of Design, Architecture, Art and Planning, the team is developing a mental health symbolic language, a system to visualize a patient’s likelihood of developing anxiety, depression or suicidal thoughts. “When we provide results, they need to be understandable by clinicians,” says Pestian.
This approach will help provide earlier recognition of clinically relevant mental health conditions in children. With the AI, Glauser says, “we can single out high-risk young people much earlier than the medical system is currently identifying them.”
In collaboration with the University of Cincinnati’s College of Design, Architecture, Art and Planning, the team is developing a system to visualize a patient’s likelihood of developing anxiety or depression. Pictured, a sample illustration of these trajectory visualizations.Credit: Cincinnati Children’s Hospital Medical Center
Computationally, there remains a great deal of algorithm development, fine-tuning and data science to be done before the trajectories are ready for paediatricians to use, not to mention the challenge of introducing new technology. But Pestian is confident that he and his colleagues will make their vision a reality. “We have a great team and great organizations supporting us.”
Ultimately, the researchers predict the AI will help prevent certain mental illness altogether. “If we can find those kids whose experiential episodes put them at high risk of anxiety and depression, we can reach out to them sooner and support them. We’re headed in that direction,” says programme partner Michael Sorter, director of child and adolescent psychiatry at Cincinnati Children’s.
“Right now,” Sorter adds, “we really have to nail down the science.” When they do, the programme has a chance to set the mental health crisis among children on a new, more positive trajectory.