EXPLAINABLE AI FRAMEWORK FOR EARLY AUTISM SPECTRUM DISORDER DETECTION: INTEGRATING ENSEMBLE LEARNING WITH CLINICAL INTERPRETABILITY

Abstract

Autism Spectrum Disorder (ASD) diagnosis is often delayed due to subjective assessments and heterogeneous symptoms. Current screening methods lack objectivity and scalability, highlighting the need for computational approaches that balance predictive accuracy with interpretability. To develop and validate a machine learning framework for ASD prediction by integrating ensemble learning, Synthetic Minority Oversampling Technique (SMOTE), and explainable artificial intelligence (XAI) to address class imbalance and ensure diagnostic transparency. Four UCI datasets comprising 3,743 instances across children, adolescents, young adults, and adults with 18 demographic, familial, and AQ-10 features were analysed. SMOTE balanced training data (1,593 per class). Nine classifiers and two ensembles (Voting, Bagging) were evaluated using accuracy, precision, recall, F1-score, and AUC with five-fold cross-validation. Model interpretability was achieved through SHapley Additive exPlanations (SHAP). CatBoost achieved the highest performance (AUC 0.9987, accuracy 0.9853) with balanced precision and recall. XGBoost (AUC 0.9986) and Voting Ensemble (AUC 0.9979) also performed strongly. Cross-validation confirmed stability (SD 0.0023). SHAP highlighted ethnicity (14.18%), age (11.71%), family ASD history (6.97%), and AQ items (A7, A9, A1, A6, A8, A2) as key predictors. The framework combines exceptional predictive accuracy (AUC > 0.99) with transparent interpretability. SHAP-based insights align with clinical knowledge, while robust validation demonstrates strong generalisation, positioning this approach as a promising tool for early ASD screening. This study integrates ensemble learning, class balancing, and XAI into a scalable, objective ASD screening tool that preserves clinical interpretability. With ~99% sensitivity, it reduces missed cases and—by providing transparent, case-level explanations—can accelerate referrals and improve access to early...