Back in February, we brought you news of a deep-learning algorithm able to predict autism in two-year-olds based on structural brain changes beginning at six months of age.
Now, the same group at the University of North Carolina has again applied machine learning to the goal of predicting autism, with equally impressive results. This time, instead of structural changes, they were able to detect changes in brain function of six-month-olds that predicted if the children would later develop autism.
The study is notable because there were no false positives—that is, all the children predicted to develop autism did.
There were a few misses, however. Of 59 6-month-old infants at high-risk for autism—meaning they had at least one sibling with autism—the algorithm correctly predicted 9 of 11 who later received a positive diagnosis.
By combining this functional analysis with the earlier structural results, it is very possible one could create a highly sensitive and accurate early diagnostic test for autism, says first author Robert Emerson of the Carolina Institute for Developmental Disabilities at UNC. And AI is going to be key to making that happen, he adds.
“It’s going to be really important to use machine learning in the future to pull all these pieces of information together,” says Emerson. In addition to brain scan data, researchers could gather behavioral results, environmental exposures, and more. Once that is done, “we’re going to have a very good shot at really nailing this early prediction.” The paper is published today in the journal Science Translational Medicine.
The team, led by UNC’s Joseph Piven and John Pruett at the Washington University School of Medicine, scanned the brains of infants while they slept. The children were again scanned at age two and completed behavioral and clinical assessments. Each functional connectivity MRI (fcMRI) scan measured the activity of 26,335 brain connections among 230 brain regions.
Using that data, a machine-learning algorithm analyzed how the activity of each piece of the brain was synchronized with other pieces of the brain. The team focused on brain regions associated with key features of autism, such as language skills and repetitive behaviors.
Brain scan signatures that predicted later autism diagnosis in infantsGraphic: R.W. Emerson et al., Science Translational Medicine (2017)
The algorithm was able to pick out functional connections that corresponded with autistic behaviors. In some cases, brain connections of children diagnosed with autism were highly synchronized; in others, the connections were less synchronized. But overall, there were functional signatures in the brain that the computer was able to detect.
Finally, the researchers applied the algorithm individually to each child at age 6 months. It was able to predict 9 of 11 children who would go onto be diagnosed with autism. “It did miss two kids, so it might be that their particular type of autism wasn’t represented in the behavioral profile we used to pick out the brain connections,” says Emerson.
Overall, these findings and the earlier results demonstrate that brain changes in autism occur during early development, prior to the onset of symptoms, says Piven. And if they can be replicated and confirmed in a larger group of children, doctors could begin to identify kids at risk of autism very early in life and enroll them in clinical trials to discover therapies for the disorder.
The study is the most recent of four papers resulting from the Infant Brain Imaging Study, a U.S. National Institutes of Health–funded study of early brain development in autism. In addition to the paper on structural changes, the team has published results linking increased cerebrospinal fluid with autism diagnoses and identified brain networks involved in attention deficits associated with autism.
Megan is an award-winning freelance journalist based in Boston, Massachusetts, specializing in the life sciences and biotechnology. She was previously a health columnist for the Boston Globe and has contributed to Newsweek, Scientific American, and Nature, among others. She is the co-author of a college biology textbook, “Biology Now,” published by W.W. Norton. Megan received an M.S. from the Graduate Program in Science Writing at the Massachusetts Institute of Technology, a B.A. at Boston College, and worked as an educator at the Museum of Science, Boston.