KAIST develops AI that analyzes mouse behavior to detect autism

KAIST researchers developed an AI model called BehaVERT that interprets mouse movement patterns like language, successfully identifying autism-related behavioral traits in mice without prior biological knowledge. The model outperformed existing benchmarks across five international standards.

A team at the Korea Advanced Institute of Science and Technology KAIST has developed an artificial intelligence model that interprets mouse movement patterns the way language models interpret text, and used it to independently identify behavioral traits linked to autism in mice, the university said Wednesday. The model, called BehaVERT, was developed by a team led by professor Kim Dae-soo in KAIST's Department of Brain and Cognitive Sciences. It converts the skeletal movement of the nose, ears, spine, limbs and tail into "tokens," the basic units language models use to process words, then feeds them into a BERT-based transformer, a type of AI architecture originally built for natural language processing. Without being given any prior biological knowledge, the model successfully identified core social and behavior deficits in mice bred to model autism, the researchers said. It surpassed existing benchmarks across five international standards covering social interaction, multianimal behavior, 3D motion analysis and autism-related behavior analysis. In tests distinguishing autism-model mic