Depression can be hard to diagnose, let alone to predict in advance. This machine learning system could highlight individuals who may be vulnerable to depression -- based only on their MRI scans.
Depression can be a crippling disorder, affecting upward of 15 million American adults and representing the leading cause of disability for people between 15-44.
New research coming out of the University of Texas at Austin could make it easier to diagnose, however — or even to highlight individuals who could be vulnerable to depression prior to its onset.
“There’s a whole lot of literature that’s emerging in the field of cognitive neuroscience and psychiatry that looks at using in vivo brain imaging techniques in humans to examine differences that might be associated with mental disorders,” David Schnyer, a cognitive neuroscientist and professor of psychology at the University of Texas at Austin, told Digital Trends. “The issue with a lot of that work is that it’s largely descriptive; it doesn’t tell us how to look at differences in brain scans in a way that gives us predictive power. That would mean telling us who has a mental disorder, or even predicting who is vulnerable to one. Our work aims to address that.”
The work Schnyer refers to involved training a machine learning artificial intelligence system to spot patterns in data to make predictions about possible depression and anxiety.
It’s a potentially exciting advance that could be useful in answer a long-term question: Whether it is possible to predict mental disorders by exploring the relationship between brain structure as seen in neuroimaging data and brain function. In their proof-of-concept study, the machine learning system was able to predict with 75 percent accuracy whether an individual had a major depressive disorder based on magnetic resonance imaging (MRI) scans. This was achieved by analyzing brain data from 52 individuals seeking treatment for depression, as well as 45 healthy control participants.
The plan next is to try and further improve the algorithm using the university’s Stampede 2 supercomputer, which will allow the team to expand their study to include data from hundreds of volunteers from the Austin area, who have been diagnosed with depression, anxiety, or other related conditions.
As Schnyer points out, another key aspect of machine learning is that the researchers can analyze multiple data types at once. “In other words, you can create algorithms with multiple kernels — so you can feed it with MRI data, genomics data, and typical behavioral assessment data,” he said. “When you do that, the hope is that you can get a very highly accurate classification algorithm.”
The results are tentatively promising, although it may be a while before your physician is using this as a standard clinical appraisal tool.
“The idea that we can utilize this data as a diagnostic tool is the hope,” Schnyer said. “Is it going to roll out tomorrow? Probably not, but I think it’s more promising than a lot of the other more descriptive work has been.”