A Novel Feature Optimization Approach for Accurate Autism Spectrum Disorder Prediction

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition characterized by challenges in communication, social interaction, and behavior. Timely detection is critical for early intervention, yet traditional diagnostic practices are often subjective, time-consuming, and prone to inaccuracies. This study addresses the limitations of existing Machine Learning (ML) and Deep Learning (DL) models in the prediction of ASD, particularly suboptimal performance caused by irrelevant or redundant features. The primary objective was to develop a robust and accurate ASD prediction framework using a novel feature optimization approach called CNN-ET-XGB. The proposed model integrates Convolutional Neural Networks (CNNs) to extract high-level abstract features from behavioral questionnaire data, Extra Trees (ET) to select the most relevant and discriminative features, and Extreme Gradient Boosting (XGB) for final classification. The model was evaluated on the UCI ASD children dataset, with a 50:50 train-test split, achieving 99.992% accuracy, outperforming existing models such as Random Forest, AlexNet CNN, and other approaches. The CNN-ET-XGB framework demonstrates significant potential for real-world applicability in clinical pre-screening tools and early ASD detection systems. Its layered feature optimization strategy enhances both accuracy and efficiency, providing a solution for assisting healthcare professionals in early ASD prediction.