Discriminating Bipolar Disorder from Major Depression using Whole-Brain Functional Connectivity: a Feature Selection Analysis with SVM-FoBa Algorithm

It is known that both bipolar disorder (BD) and major depressive disorder (MDD) can manifest depressive symptoms, especially in the early phase of illness. Therefore, discriminating BD from MDD is a major clinical challenge due to the absence of biomarkers. Feature selection is especially important in neuroimaging applications, yet high feature dimensions, low sample size and poor model understanding present huge challenges. Here we developed an advanced feature selection algorithm, “SVM-FoBa”, which enables adaptive selection of informative feature subsets from high dimensional resting-state functional connectives (rsFC) data. By comparing SVM-FoBa with conventional feature selection methods on several public biomedical data sets, the proposed method was proven to be increasingly superior as the feature dimension became high. When applying SVM-FoBa to brain data, with 38 significant rsFCs chosen from 6670 in total, an 88 % classification accuracy between BD and MDD was achieved using leave-one-out cross-validation. Further, by conducting weight analysis, the most discriminative FCs were revealed, providing which adds to our understanding of functional deficits and may serve as potential biomarkers for mood disorders.