Distinguishing between bipolar depression and unipolar depression based on the reward circuit activities and clinical characteristics: A machine learning analysis.

Background: Differentiating bipolar depression (BD) from unipolar depression (UD) is a major clinical challenge. Identifying the potential classifying biomarkers between these two diseases is vital to optimize personalized management of depressed individuals.

Methods: Here, we aimed to integrate neuroimaging and clinical data with machine learning method to classify BD and UD at the individual level. Data were collected from 31 healthy controls (HC group) and 80 depressive patients with an average follow-up period of 7.51 years. Of these patients, 32 got diagnosis conversion from major depressive disorder (MDD) to BD (BD group) and 48 remain persistent diagnosis of MDD (MDD group). Using graph theory and functional connectivity (FC) analysis, we investigated the differences in reward circuit properties among three groups. Then we applied a support vector machine and leave-one-out cross-validation methods to classify BD and UD patients based on neuroimaging and clinical data.

Results: Compared with MDD and HC, BD showed decreased degree centrality of right mediodorsal thalamus (MD) and nodal efficiency (NE) of left ventral pallidum. Compared with BD and HC, MDD showed decreased NE of right MD and increased FC between right MD and bilateral dorsolateral prefrontal cortex and left ventromedial prefrontal cortex. Notably, the classifier obtained high classification accuracies (87.50 %) distinguishing BD and UD patients based on reward circuit properties and clinical features.

Limitations: The classifying model requires out-of-sample replication analysis.

Conclusion: The reward circuit dysfunction can not only provide additional information to assist clinical differential diagnosis, but also in turn informed treatment decision of depressive patients.