Differences in resting-state functional connectivity between depressed bipolar and major depressive disorder patients: A machine learning study

Nearly 60 % of individuals with bipolar disorder (BD) are initially classified as major depressive disorder (MDD) patients, resulting in inappropriate drug treatment. Identifying reliable biomarkers for the differential diagnosis between MDD and BD patients may allow to define the best treatment option since the early phases. In this study, we deployed machine learning predictive models to classify 62 MDD and 63 BD patients with a current depressive episode from resting functional neuroimaging feature (rs-fMRI), including fractional amplitude of low-frequency fluctuations, regional homogeneity, atlas-based connectivity across 434 regions of interest, seed-based connectivity maps for 44 seeds, and 14 dual regression components. Models were also compared to 76 healthy controls. Only the model trained on seed-based connectivity reached the statistical significance in permutation test reaching the highest classification performance (69.36 % of accuracy for BD and 63.08 % for MDD). Seed-based connectivity also reached the best performance in identifying MDD (78.33 %) and BD (71.67 %) relative to controls. Connectivity patterns in key brain regions of the reward and aversion systems appeared crucial in differentiating the disorders, possibly identifying distinct clinical phenotypes of disorders, beyond the depressive ongoing episode.