Cross-continental team mines autism biomarkers from fMRI scans

Clinical and neuroscience researchers in Japan and Providence, R.I., have put their heads together and come up with an algorithmic classifier that can distinguish between autistic and non-autistic brains as imaged with fMRI.

The achievement may point the way to the invention of a diagnostic tool, according a study report published April 14 in Nature Communications.

Noriaki Yahata, PhD, University of Tokyo, Yuka Sasaki, PhD, Brown University, and colleagues successfully tested the classifier with 181 adult volunteers at three sites in Japan (74 with autism diagnoses, 107 without), then validated it in a group of 88 American adults (44 with, 44 without) at seven sites.

The classifier, which combines two machine-learning algorithms, allowed the researchers to analyze differences in 16 functional brain connections that, they found, were indicative of autism.

The classifier averaged 85 percent accuracy among the Japanese participants and 75 percent accuracy among the Americans.

Additionally, the team showed the classifier can be impressively predictive of the autistic participants’ degree of impairment as scored by clinicians in the Autism Diagnostic Observation Schedule (ADOS).

In the ADOS communications component, the classifier predicted scores with a statistically significant correlation of 0.44.

The autistic participants had no intellectual disability, the authors note.

“These results indicate that although we developed a highly reliable classifier by only using the training data obtained from Japan, it is sufficiently universal to classify [autism] in the U.S.A. validation cohort,” they write in their discussion.

In an article from the news office at Brown, Sasaki says the classifier represents a significant breakthrough even though it is not yet ready for clinical primetime.

“Eighty percent accuracy may not be useful in the real world,” he says.

Still, he notes, this study is the first to successfully apply a classifier “to a totally different cohort. There have been numerous attempts before. We finally overcame the problem.”