Research on Identification of Children's Autism Spectrum Disorders based on EEG Time-Frequency Domain Feature Fusion and XGBoost

Abstract

Accurate early diagnosis of Autism Spectrum Disorder (ASD) lack’s objective and quantitative biomarkers. Existing EEG-based ASD identification studies are plagued by insufficient samples (usually <50 cases), single feature extraction and other flaws, which severely degrade the model's generalization ability and clinical applicability. Based on private resting-state EEG data of 603 children, this study constructed the MIFE-XGBoost intelligent identification model integrating time-frequency domain features and a two-stage optimization strategy. Raw EEG signals were standardized preprocessed, then time-domain nonlinear features and periodogram-based frequency-domain power features were extracted and fused; a "mutual information feature screening - EEG channel weight fusion" strategy was designed to optimize the feature set for binary classification of ASD and Typically Developing (TD) children. Experimental results show the model achieved an overall accuracy of 92.3±0.7%, an ASD recall rate of 91.8±0.9% and an F1-score of 91.4±0.8%, with robust performance and good clinical deployment potential. This study improved model generalization via large-sample data, identified effective EEG time-frequency features for distinguishing ASD and TD children, providing an efficient method for early pediatric ASD screening and new technical reference for neurodevelopmental disorder identification based on EEG signals.