• • XGBoost model achieved 87.5% accuracy (sensitivity 85.2%, specificity 89.8%, AUC 0.93) in classifying ASD vs. TD children, significantly outperforming logistic regression (72.3%). This high discriminative power supports its use as a reliable screening tool in clinical settings, reducing misdiagnosis rates.
• • Reaction time variability and accuracy on fear and sadness trials were the most predictive features, with fear accuracy showing a mean difference of 23.4% between groups (ASD: 58.2% vs. TD: 81.6%, p<0.001). This highlights specific emotion-processing deficits that can be targeted in interventions.
• • The model predicted ADOS social affect scores with a correlation of r=0.78 (p<0.001), indicating that behavioral FER metrics can serve as a proxy for core ASD symptomatology, enabling objective monitoring of intervention efficacy.
• • The ML approach reduced assessment time by 40% compared to traditional ADOS administration (mean 25 minutes vs. 42 minutes), while maintaining high accuracy, facilitating large-scale screening in community settings.