Researchers at Seoul National University Hospital (SNUH) have developed an AI platform using objective and biological indicators based on magnetic resonance imaging (MRI) brain images to diagnose autism spectrum disorder (ASD) rapidly and accurately.

ASD is a neurodevelopmental disorder that affects about 1 to 2 percent of children. The illness leads to difficulties forming social relationships, problems in emotional interaction, repetitive behaviors, and limited interests.

The cause of the disease is known to be due to the interaction of genetic and environmental factors. However, reports have been increasing recently that abnormal development of the structure and function of the social brain also causes the disease.

Hospitals have diagnosed ASD through symptom evaluation after observing abnormal behaviors or expressions during development. Although this diagnostic method has a high degree of agreement among experts, there is a limit due to the observer's subjectivity to intervene, making it difficult to identify the causal relationship. Accordingly, the need for research to confirm the possibility of a diagnosis of ASD based on objective and biological indicators has emerged.

The team, led by Professor Kim Bung-nyeon of the Department of Pediatrics, evaluated the diagnostic discrimination ability of 58 patients with ASD and 48 people in control groups through an MRI brain image-based machine learning AI algorithm. The participants were three to six years old, and the team only included low-functioning patients in the autism group.

Researchers built the machine learning algorithm as a classifier by applying machine learning, such as the random forest. As parameters for classification, the researchers used T1-weighted MRI images, diffusion tensor images, and multiple MRIs. Afterward, they evaluated the ability to distinguish between the autism group through a machine learning AI algorithm.

As a result, the multiple MRI model showed high diagnostic discrimination ability with 88.8 percent accuracy, 93 percent sensitivity, and 83.8 percent specificity.

Notably, the accuracy of the multiple MRI model improved by 10 percentage points compared to T1-weighted MRI and diffusion tensor imaging.

Also, the team confirmed that the most important imaging parameters for diagnosing autism spectrum disorder were the cortical thickness of the occipital lobe, diffusion of the cerebellar horn, and posterior cingulate gyrus connectivity.

"This study confirmed the use of multiple MRIs through machine learning is useful in diagnosing patients with ASD in infants and young children with severe developmental delay based on biomarkers," Professor Kim said. "We expect the accuracy of autism diagnosis to improve if we add functional brain imaging data and supplement it to multiple MRIs through additional research in the future."

Journal of Autism and Developmental Disorders has published the study results in its latest issue.