Ensemble Deep Neural Networks for Reliable Autism Diagnosis in Children: A Performance and Interpretability Study

Detecting autism spectrum disorder (ASD) in children poses challenges at earlier stages, considering the absence of common screening tools and subtle behavioral patterns. Existing machine learning (ML) models suffer from generalization, class imbalance, and interpretability issues, particularly when applied to small or structured datasets. Many studies also utilize single-model diagnostic systems, which are unlikely to represent complex patterns in diverse datasets. This paper aims to propose an ensemble of deep learning (DL) models for detecting ASD and attempts to resolve the aforementioned problems using structured screening data. Our ensemble comprises a deep neural network, a one-dimensional convolutional neural network, and an autoencoder-based classifier. We perform the model evaluation through stratified five-fold cross-validation to maintain the subsets’ consistencies. As such, the method achieved an average accuracy of 98.4%, an F1 score of 98.4%, an area under the curve of 0.9997, and a Matthews correlation coefficient of 0.969. SHapley Additive exPlanations (SHAP) values and t-distributed stochastic neighbor embedding (t-SNE) plots elucidate the model’s decisions and illustrate class distinguishability. Overall, the results highlight the capability of DL ensembles as a robust solution for early ASD detection in children while maintaining explainability. Despite the promising results, the model’s generalizability is constrained by the small dataset size and its reliance on static, questionnaire-based data. Broader clinical validation is required to support its real-world applicability.