MАCHINE LEARNING АLGORITHM FOR EARLY DETECTION OF АUTISM SPECTRUM DISORDERS IN CHILDREN BASED ON MULTIMODAL ANALYSIS OF EYE MOVEMENTS АND FACIAL EXPRESSIONS

Abstract

This paper presents an algorithm for extracting and analyzing informative multimodal features of oculomotor activity and facial expressions for the early detection of autism spectrum disorders (ASD). The proposed approach integrates eye-tracking and computer vision technologies to capture synchronized behavioral signals within a standardized experimental protocol based on segmented scenarios (F1-F6). A multimodal dataset consisting of temporal signals and aggregated features was constructed to represent visual attention, gaze–object interaction, and facial expression dynamics. The extracted features are transformed into a structured feature space and used to solve a binary classification problem distinguishing ASD and typically developing (TD) children using machine learning methods. Experimental results demonstrate that the integration of multimodal features significantly improves classification performance. Logistic regression achieved the best results (ROC-AUC = 0.976, PR-AUC = 0.978), outperforming SVC_RBF and Random Forest models. The obtained results confirm that multimodal behavioral features provide a reliable and interpretable representation of ASD-related patterns. The proposed method can serve as a foundation for intelligent decision support systems for the early diagnosis of ASD.