Detecting children with autism spectrum disorder based on script-centric behavior understanding with emotional enhancement

Early diagnosis of autism spectrum disorder (ASD) is critical but often hindered by clinician scarcity and subjective interpretation of social behaviors. While machine learning can assist, it typically demands large, labeled datasets and lacks interpretability. To address these limitations, we introduce a zero-shot and few-shot framework for ASD detection based on Script-Centric Behavioral Understanding (SCBU) with emotional enhancement. Our method transforms multi-view audio-visual recordings of standardized social interactions into structured, time-stamped behavioral scripts. These scripts encapsulate gaze, gestures, head pose, and speech, augmented by an emotion textualization module that captures dynamic affective changes. A domain prompt module integrates clinical knowledge, and the resulting scripts are analyzed by Large Language Models (LLMs) to infer ASD status and generate diagnostic rationales. We further propose SCBU-Agents, a multi-LLM collaboration strategy that achieves consensus through re-analysis and discussion. Evaluated on a clinical dataset of 95 children, our framework achieved an F1-score of 95.24 %, outperforming both behavioral signal processing baselines and human raters while preserving interpretability. This work demonstrates that textualizing behavior and leveraging LLM reasoning can provide accurate, interpretable, and privacy-preserving clinical decision support for ASD screening.

Duke Scholars

Author Ming Li DKU Faculty