AUTISM SPECTRUM DISORDER DETECTION USING GRAPH-BASED NEURAL NETWORKS APPROACH

Abstract

Autism Spectrum Disorder (ASD) is a neuro-developmental condition characterized by challenges in social interaction and communication. Early and accurate diagnosis is crucial but remains challenging due to complex behavioral patterns and imbalanced datasets. Existing models lack robustness in handling data imbalance and often overlook the relational structure between features. This creates a significant research gap in developing a reliable and generalizable model for ASD prediction. The problem addressed in this study is the ineffective classification of ASD cases caused by class imbalance and feature selection. The objective is to design a model that improves diagnostic accuracy while handling these limitations. To achieve this, an ensemble deep learning model combining Graph Attention Network (GAT) and Artificial Neural Network (ANN), called GAT-ANN, is proposed. The methodology involves feature selection using Graph-based Feature Selection (GBFS), class imbalance handling through GAT-based class imbalance handler and GAN-based data augmentation, and classification using the GAT-ANN model. Results show that the model achieved 99.98% and 99.99% accuracy for children and adult datasets. These findings demonstrate good performance of GAT-ANN over traditional models. In conclusion, the GAT-ANN model provides a reliable framework for ASD prediction.