Deep learning algorithm successfully predicts autism in 6-month-olds

An artificial intelligence (AI) algorithm predicted whether or not a 6-month-old child would develop autism with 96 percent accuracy in a study on a group of almost 60 infants. Early-warning screening like this could help identify at-risk children even before symptoms arise, opening up treatment pathways to mitigate the effects of autism, according to authors on the study published in Science Translational Medicine.

The researchers scanned more than 230 regions of the brain, examining if paired areas known as functional connections were in sync or not. The children returned to the lab at 2 years of age for a behavior assessment, where the researchers measured social interactions, communication skills and motor development to make an autism diagnosis.

Once they had the before-and-after data, researchers fed the data into a deep learning program—but with a catch. They used data from 58 out of the 59 infants in the study to train the program to spot associations, then tested it on the 59th infant's scans.

The algorithm correctly predicted normal development in 48 infants and predicted autism in nine out of the remaining 11—an extremely accurate result—said lead author Robert Emerson, a former postdoctoral fellow at the University of North Carolina (UNC) School of Medicine.

Previous studies have shown similar predictive value but with scans at 6 months and another at 1 year. Doing away with the later scan—therefore making the autism determination earlier—could have big implications for treatment, according to Joseph Piven PhD, Professor of Psychiatry at the UNC School of Medicine.