NIMH Working on Harnessing AI for Psychiatry

Artificial intelligence (AI) and computational biology may one day change the way psychiatrists care for patients with depression, said National Institute of Mental Health (NIMH) Director Joshua Gordon, M.D., Ph.D., during a presentation at the APA Spring Highlights Meeting this past weekend.

“You are all aware of the revolution in data science and artificial intelligence that has been changing the way we think about science and society in general,” Gordon said. “We want to take that revolution and harness it for psychiatry.”

Gordon said that researchers are examining the ways in which AI might use information contained in large collections of patient electronic health records to identify people at risk of suicide. As Gordon has noted in previous APA lectures, suicide prevention is a priority research area under his watch, given the continued rise in suicide rates in the United States over the past two decades. “Identifying people at risk of suicide is one of the key roles psychiatrists play in health care,” he said, adding that it is also one of the more difficult tasks for the field.

As reported in one recent NIMH-funded study, a combination of expanded screening of patients admitted to the emergency department coupled with telephone-based follow-up of patients who screen positive for suicide risk can reduce future attempts by 20%. “Can we use modern computational tools to improve that even further?” he asked. Preliminary work studying veterans suggests that AI may be able to predict those veterans most likely to attempt suicide over a 12-month period, he said, adding that NIMH is seeking more grant applications related to the prediction of suicide risk.

Gordon also described how computational phenotyping might lead to greater understanding of psychiatric disorders down the road. As Gordon explained, computational phenotyping involves trying to mathematically define symptoms of a disorder. Asking where in the brain happiness is controlled, for example, is not a feasible way to understand depression, he noted. “But we can ask questions like, ‘Where in the brain are these different behavioral components that together make up happiness?’”

He highlighted one study that used brain imaging data of people playing a game of chance on their phones to calculate reward prediction error. Basically, when people are about to decide something involving reward/loss, they have some expectations of how they will react to winning or losing, which guides their decision. Normally the outcomes and expectations match, but in many people with depression, reward prediction is off (rewards are not as pleasurable as thought, which lowers a desire for future rewarding decisions, for example). This study found that reward prediction in the participants could be localized to two discrete brain regions: the ventral striatum and medial prefrontal cortex.

Other brain imaging studies have found these same two regions show reduced activity in many depressed patients. Gordon noted that linking specific brain regions with particular behavioral traits might help with precision medicine efforts. Based on a patient’s symptom profile, physicians might be able to determine which brain regions are affected and then identify the optimal treatment. ■