Machine Learning-Based Early Detection Of Autism Using Multimodal Conversational Features

Abstract

Early and reliable screening for autism spectrum disorder (ASD) remains challenging in low-resource and high-variance conversational settings. This paper presents an end-to-end multimodal screening system that analyzes child-caregiver interaction data from audio recordings, CHAT-format transcripts, and text inputs to estimate ASD likelihood and provide clinician-facing explanations. The system integrates three feature families: pragmatic-conversational, acoustic-prosodic, and syntactic-semantic, supporting component-wise classification and late-fusion strategies with modality-aware weighting. Beyond prediction, the platform provides transcript-level behavioral annotations, global and local feature attributions, and counterfactual what-if analysis. Experiments on cross-validated ASDBank data show multimodal fusion achieving 87.2% accuracy (ROC-AUC 0.92), outperforming unimodal baselines by 2-4%.