Feature-engineered stacking and AutoGluon framework for deep learning-based autism detection

Abstract

Autism spectrum disorder (ASD) is a brain development disorder that distracts a person’s perceptions and social interactions with others. Current studies emphasize conventional machine learning and typical ensemble methods and little focus have been placed on merging automated machine learning (AutoML) systems with stacking ensemble structures. To bridge this gap, this study evaluates stacking ensembles with an artificial neural network (ANN)-based meta-classifier and the AutoGluon framework. For enhancing model stability and generalizability across various ASD datasets, systematic RandomizedSearchCV along with 10-fold stratified cross-validation is utilized. The stacking model outperformed the AutoGluon framework and all individual classifiers on dataset-I, achieving 99.50% accuracy, 99.53% precision, 99.53% recall, and a 99.97% receiver operating characteristic-area under the curve (ROC-AUC). With 97.28% accuracy, precision 96.33%, 95.38 F1-score, and a 99.76% ROC-AUC on dataset-II, the model demonstrated great generalization and continuously outperformed traditional ensemble and boosting techniques. Stacking model perform well on both datasets and maintain consistence, whereas AutoGluon give competitive performance on dataset-1 and it exhibited a notable decline on dataset-II. K-fold internal and external cross-validation were used to access model’s generalization and consistency. The outcomes of stacking ensemble model indicate that model offers higher accuracy, greater resilience, better consistency across difference datasets and offers a robust approach to early ASD prediction.