AI enables precise autism diagnosis, opening path to personalized treatment

Researchers have developed an AI-powered technology that analyzes brain changes associated with autism spectrum disorder (ASD) and predicts both diagnosis and treatment progress. The advance moves beyond conventional diagnostic methods that rely on observing symptoms and behavior, offering an objective way to assess functional brain networks.

An international research team including Han Gyeong-rim of the Korea Institute of Science and Technology's Brain Science Institute announced Tuesday that it had developed an interpretable AI framework combining functional magnetic resonance imaging (fMRI) with AI, in collaboration with the University of Melbourne and Seoul National University Hospital. The findings will be presented in the "AI for Science" track at ACM SIGKDD 2026, held at the Jeju International Convention Center.

Using fMRI analysis, the team found that functional brain connectivity in children with autism is lower than in typically developing children. The reduction was particularly pronounced around the inferior parietal lobule, which integrates sensory information, and the thalamus, which relays sensory signals.

The key innovation is that the AI does not simply determine whether a person has autism — it can also explain its reasoning through observed changes in brain networks. The team developed a "Generative Manifold Auditing" (GMA) framework, training an AI on the brain connectivity patterns of typically developing individuals, then blocking connections in specific regions to analyze the AI's responses. A "sensitivity gap" metric verified that the AI was capturing genuine structural changes in brain networks rather than mere information loss.

Validation using ABIDE, a large-scale autism brain imaging dataset, and a clinical cohort from Seoul National University Hospital showed that a generative AI model called a Denoising Autoencoder distinguished structural changes in brain networks more accurately than comparable models.

Notably, even among patients whose clinical symptoms had improved, changes in functional brain networks varied from person to person. The team said this points to the potential of using AI to analyze individual brain changes that are difficult to detect through symptoms alone.

"This demonstrates the possibility of objectively analyzing brain changes that differ from patient to patient," Han said. "We will continue this work toward developing biomarkers that can be used for early diagnosis and predicting treatment progress."