Researchers have built an AI system that screens for autism by reading brain scans and facial expressions together, rather than relying on images alone.
The system, called UniAR, tackles a real bottleneck: there isn't much labeled clinical text available to train on, so most autism-detection AI has leaned on visual data by itself. UniAR's fix is to have a large multimodal model generate its own descriptions of the scans at the word, phrase, and sentence level, then match those descriptions to visual patterns using a mixture-of-experts alignment module. Tested on four benchmarks spanning brain MRI and facial-expression data, it hit 75.9% average accuracy on MRI and 91.6% on facial scans, edging out prior state-of-the-art methods by 1.5 and 1.2 percentage points respectively.
Autism screening today leans heavily on behavioral observation and clinician judgment, which is slow and unevenly available depending on where you live. A tool that can read imaging data with some interpretability, and that doesn't need mountains of annotated clinical reports to get there, could make earlier screening cheaper to run at scale. That matters because earlier intervention is consistently linked to better developmental outcomes.
The accuracy gains here are incremental, not a leap, and this is still a research paper, not a clinical tool anyone can order today - worth remembering before this gets rebranded as an AI 'test' for autism.