New AI Screens Kids for Autism by Watching Videos

Clinical management of autism spectrum disorder hits a bottleneck at early screening, the paper says, mainly because trained specialists are scarce and conventional assessment tools are subjective. Enter ASDchat: a multimodal large language model built for evidence-based ASD screening that takes video, audio, and dialogue as input.

It runs on a dual-branch architecture. The decision branch generates screening probabilities; the evidence branch generates traceable, timestamped behavioral evidence aligned with standardized clinical criteria (ADOS-2) — so the model doesn't just flag, it points to what it saw.

The model was trained and evaluated on 1,035 participants from 27 sites in China, spanning typically developing children, children with ASD, and children with other disorders. For ASD versus typically developing children, ASDchat reached an area under the receiver operating characteristic curve (AUC) of 0.953 ± 0.021 — AUC being a metric where 0.5 is chance level, not a percentage of correct predictions. On 9 held-out sites that weren't used for training, the mean AUC was 0.932.

There's more: unsupervised clustering of the behavioral dimensions split the ASD cases into six subtypes with different phenotypic profiles, and ASDchat suggests an intervention for each subtype.

The paper's conclusion: ASDchat provides a feasible path for large-scale, evidence-based early ASD screening in clinical practice.