Original title: A multimodal large language model for evidence-based autism spectrum disorder screening
Jun Chen, Qi Zhao, Yunliang Jiang, Shuqin Cao +7 · September 2026
Doctors struggle to screen young children for autism spectrum disorder because specialists are rare and standard tests can be subjective. To solve this, researchers built ASDchat, an artificial intelligence model that looks at video, listens to audio, and reads dialogue. It works using two pathways at once. One pathway calculates the chance that a child has autism. The other pathway finds the proof, picking out specific moments in the recording that show behaviors matching standard clinical criteria. When tested, ASDchat successfully spotted autism, scoring high accuracy even at clinics it had never seen during training. It also grouped children with autism into six behavioral subtypes and suggested a specific intervention for each group.
A teacher is trying to figure out which students are struggling. Relying on a single test score often misses important details like who asks confused questions or looks lost during a lesson. In this paper, the model takes in video, audio, and dialogue—just like a teacher observing class—to produce both a likelihood of a condition and specific, time-stamped observations that explain the decision. It even groups students with similar profiles and suggests different teaching strategies for each group.
Early autism screening is limited by a shortage of trained specialists and reliance on subjective clinical judgment. ASDchat could make screening accessible to more families by using video, audio, and dialogue to provide objective, traceable assessments. It also identifies six subtypes with tailored intervention suggestions, potentially guiding personalized treatment.
The model was trained and tested only on Chinese participants, so its performance in other populations is unknown. It relies on the ADOS-2 criteria, a clinical standard, but the evidence generation has not been clinically validated for real-world use. The intervention suggestions for subtypes are based on unsupervised clustering that may not correspond to clinically meaningful categories.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
A multimodal large language model for evidence-based autism spectrum disorder screening — Jun Chen, Qi Zhao, Yunliang Jiang et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.