Combat harmonization of rs-fMRI improves classifier performance in patients with autism spectrum disorder: a graph theoretical approach

Abstract

Resting state functional connectivity is identified to reflect the intrinsic organization of the brain’s cognitive characteristics. One key property of such biological networks is modularity. In this paper, an open dataset ABIDE of which a total of 45 ASD (Autism Spectrum Disorder) and 53 TD (Typical Development) participants from KKI and PITT sites was considered. Graph Theoretic measures involving Degree Centrality, Clustering Coefficient and Modularity were calculated at different thresholds of network connectivity. Linear SVM classifiers were used to distinguish between ASD and TD. The classifier was applied to each of these graph measures at individual threshold levels.  Further, ComBat harmonization across both sites was performed to which the above classification model was applied. Experimental results show that participants from KKI sites showed a difference in Newman modularity at 0.3 threshold with 76.1% accuracy with specificity and sensitivity of 73.08% and 81.25% respectively. Although the accuracy of the site-specific TD vs ASD classifiers were comparable, the sensitivity of the KKI classifier was less, which improved post harmonization. Similar performance was observed in degree centrality and clustering coefficient measures. Therefore, the harmonization process plays a significant role in classifier performance and in model selection.