Autism Spectrum Disorder Detection Using a Graph Neural Network with Self-Supervised Contrastive Learning Framework

Autism Spectrum Disorder (ASD) is a complicated neurological disorder defined by the lack in social behaviour and communication patterns. Even though various multimodal deep learning techniques have given a promising predictions, lot of present models rely on supervised learning and shows poor performance because of  limited labelled data set. This study proposes a novel Graph Neural Network with Self-Supervised Contrastive Learning (GNN-SSCL) methodology for ASD detection using the ABIDE I dataset and eye-tracking data set for efficient prediction. The proposed approach used will make use of all the brain connectivity as graphs and learns a complex and robust relationships through a contrastive training before doing the classification. The experimental results shown illustrate that the proposed method achieves 93.2% accuracy and 0.96 AUC, which may outperform the existing transformer-based architectures and CNN-based models. The proposed method also improves generalizability, reduces dependency on the labelled data, and demonstrates a very good interpretable model for real world clinical applications.