Leveraging multi-dataset stacking ensemble for early screening autism spectrum disorder in adults with explainable AI

Due to reliance on subjective clinical assessments, a lack of specialists, and the practical limitations of large-scale behavioral examination, accurate and timely screening for autism spectrum disorder (ASD) in adults continues to be difficult. By combining Random Forest (RF), XGBoost, Logistic Regression (LR), and Support Vector Machine (SVM) classifiers into a single decision architecture, this study suggests a transparent stacking ensemble learning framework for adult ASD prediction. An integrated adult ASD screening dataset of 964 samples from the UCI Autism Screening Adult dataset and the Kaggle Autism Screening on Adults dataset, which included AQ-10 behavioral responses, demographic factors, and medical-history indications, was used to create the model. Prior to merging, statistical consistency analysis was carried out to ensure robust dataset integration. This included the Kolmogorov–Smirnov (KS) test for age and Chi-square tests for categorical variables. Bias mitigation was then achieved through feature standardization, age binning, and the removal of dataset-identifying attributes. The pre-processing pipeline also featured feature scaling, IQR-based age outlier capping, One-Hot Encoding for nominal categorical variables, and missing-value imputation. SMOTE was integrated into a 5-fold stratified cross-validation pipeline to correct class imbalance without causing leakage. This ensured that only fold-specific training partitions produced synthetic minority samples. With 94.80% accuracy, 86.9% precision, 96.4% recall, and 91.4% F1-score, the optimized stacking ensemble outperformed individual base learners and showed strong suitability for adult ASD pre-screening, where high recall is clinically required. SHAP and LIME were used to give both global and local interpretability, continuously highlighting important behavioural markers linked to ASD-related social and cognitive features to improve transparency and clinical confidence. With enhanced robustness across diverse questionnaire-based datasets and great potential for implementation in community and telehealth pre-screening settings, the suggested framework provides an all-around scalable, interpretable, and affordable adult ASD screening approach.