An Explainable and Robust Machine Learning Approach for Autism Spectrum Disorder Prediction

Abstract

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition that requires early detection for timely intervention. In this study, a comprehensive machine learning framework was developed and evaluated for predicting ASD traits using a behavioral and demographic dataset comprising 1,985 records and 28 features. Eight models, including Logistic Regression, Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Decision Tree, Random Forest, Gradient Boosting, XGBoost, and LightGBM, were systematically assessed.

Experimental results demonstrated that advanced classifiers achieved superior predictive performance, with SVM attaining the highest ROC-AUC (99.90) and Random Forest yielding the highest test accuracy (97.98%). Robust analysis using calibration curves confirmed that probability estimates were well-aligned with true outcomes, while bootstrap confidence intervals validated the stability of the reported metrics. Furthermore, interpretability was incorporated through SHAP analysis, which identified speech delay, family history of ASD, anxiety disorder, and specific AQ-10 items as key predictive features. These findings highlight the potential of explainable and reliable computational models for supporting ASD screening in clinical and community settings. The proposed framework balances predictive accuracy with interpretability and reliability, addressing key barriers to the adoption of data-driven approaches in healthcare decision support.