Functional Connectivity and Graph Embedding-Based Domain Adaptation for Autism Classification from Multi-site Data

Many machine learning-based classification models for autism spectrum disorder (ASD) using neuroimaging data have been proposed. In recent developments, research has transitioned its focus to using extensive multi-site brain imaging datasets to increase the clinical applicability and statistical robustness of findings. However, the classification performance is hampered by the inherent heterogeneity of these combined datasets. This paper introduces a novel correlation-based functional connectivity method designed to extract improved Region of Interest (ROI) coupling features from the Autism Brain Imaging Data Exchange (ABIDE) dataset. We assess graph embedding domain adaptation (GEDA) to mitigate dataset heterogeneity, mapping data points from source and target domains into a common low-dimensional space while preserving their similarity relationships. We employ a novel dataset-splitting approach called the ’rectified environment’ to enhance classification accuracy. To validate our proposed model, we compared it with related works. Our result shows that the proposed model with support vector machine (SVM) has an accuracy of 78.1% and AUROC 83.9% in identifying ASD patients. Our model demonstrates a substantial improvement, increasing accuracy by 6.1% and AUROC by 5.3% compared to the maximum independence domain adaptation (MIDA) model. These findings reveal an anticorrelation in brain function and disruptions in brain connectivity between anterior and posterior brain regions in ASD.