EXPLAINABLE AND FAIRNESS-AWARE MACHINE LEARNING FOR EARLY AUTISM RISK SCREENING USING BEHAVIORAL AND DEMOGRAPHIC DATA

Abstract

Early detection of Autism Spectrum Disorder (ASD) is essential to allow interventions that can significantly improve developmental trajectories. This paper introduces an understandable and fairness, aware machine learning approach to screening the risk of ASD at the early stage, based on behavioral and demographic data. The study used a publicly available dataset of 699 adult samples, including AQ, 10 behavioral features, and demographic attributes, to conduct experiments. Five supervised models, Logistic Regression, Support Vector Machine, Random Forest, Gradient Boosting, and K-Nearest Neighbors, were tested. Logistic Regression was found to be the best model in terms of the trade-off of accuracy (96. 43%), recall of ASD, positive cases, and interpretability. The model was made transparent through Shapley Additive Explanations (SHAP), which allowed understanding both at the global and instance level, of the most relevant to clinicians’ behavioral indicators. The fairness of the model regarding gender was measured through demographic parity, equal opportunity, and disparate impact criteria, showing some moderate bias, which was reduced through the reweighing method without causing any performance drop. The ablation study showed that behavioral features are sufficient for accurate and fair ASD screening. The proposed framework is an example of a responsible clinical AI deployment approach.