Early Intervention in Autism Spectrum Disorder: Hybridizing Convolutional Neural Networks with Ensemble Methods for Enhanced Detection Using Facial Imagery

Autism Spectrum Disorder is a neurodevelopmental condition affecting social communication and behavior. Early and accurate detection is essential for improving long-term outcomes through timely interventions. However, traditional diagnostic methods rely heavily on behavioral observations, which are time-consuming, subjective, and can vary between clinicians. The objective of this study is to propose an automated, objective method for ASD detection using facial image analysis. Specifically, we present a hybrid model that integrates Convolutional Neural Networks for feature extraction with ensemble learning methods—Random Forest and Extreme Gradient Boosting for classification. EfficientNet and MobileNet were used as feature extractors, chosen for their efficiency and accuracy in image classification tasks, while RF and XGBoost were utilized to ensure robust classification. To prevent overfitting and enhance generalization, extensive data augmentation techniques were applied, along with class balancing methods to handle imbalanced data. The proposed model was evaluated on the Autism Image Data dataset and achieved an accuracy of 97.35%, with precision and recall values of 95.2% and 95.75%, respectively. These results demonstrate the model’s potential in contributing to early ASD detection, which can be an essential tool for clinicians in diagnostic decision-making.