There is a lack of objective features for the differential diagnosis of unipolar and bipolar depression, especially those that are readily available in practical settings. We investigated whether clinical features of disease course, biomarkers from complete blood count, and blood biochemical markers could accurately classify unipolar and bipolar depression using machine learning methods. This retrospective study included 1160 eligible patients (918 with unipolar depression and 242 with bipolar depression). Patient data were randomly split into training (85%) and open test (15%) sets 1000 times, and the average performance was reported. XGBoost achieved the optimal open-test performance using selected biomarkers and clinical features—AUC 0.889, sensitivity 0.831, specificity 0.839, and accuracy 0.863. The importance of features for differential diagnosis was measured using SHapley Additive exPlanations (SHAP) values. The most informative features include (1) clinical features of disease duration and age of onset, (2) biochemical markers of albumin, low density lipoprotein (LDL), and potassium, and (3) complete blood count-derived biomarkers of white blood cell count (WBC), platelet-to-lymphocyte ratio (PLR), and monocytes (MONO). Overall, onset features and hematologic biomarkers appear to be reliable information that can be readily obtained in clinical settings to facilitate the differential diagnosis of unipolar and bipolar depression.

There is a lack of objective features for the differential diagnosis of unipolar and bipolar depression, especially those that can be easily accessible in practical settings. Some studies have shown that unipolar and bipolar depression have different associations with hematologic biomarkers and clinical features such as the age of onset. However, none of them have used these features for differential diagnosis. We investigated whether biomarkers of complete blood count, blood biochemical markers and clinical features could accurately classify unipolar and bipolar depression using machine learning methods.1,160 eligible patients were included in this retrospective study (918 with unipolar depression and 242 with bipolar depression). 27 biomarkers of complete blood count,17 blood biochemical markers and 2 clinical features were investigated for the classification. Patient data was split into training (85%) and test set (15%). Using ten-fold cross validation for training, logistic regression (LR), support vector machine (SVM), random forest (RF) and Extreme Gradient Boosting (XGBoost) were compared with feature selection.We calculated the AUC, sensitivity, specificity and accuracy. The optimal performance was achieved by XGBoost using a combination of selected biomarkers of complete blood count (WBC, PLR, MONO, LYMPH, NEUT Ratio, MCHC, BASO Ratio, LYMPH Ratio), blood biochemical markers (albumin, potassium, chlorine, HCT, calcium, LDL, HDL) and clinical features (disease duration, age of onset). The optimal performances achieved on the open test set were AUC 0.889, sensitivity 0.831, specificity 0.839 and accuracy 0.863. Hematologic biomarkers and onset features seem to be reliable information that could be easily accessible in clinical settings to improve diagnostic accuracy. In addition, we further analyzed the importance of specific blood biomarkers in samples of disease durations <= 3 years and > 3 years. WBC and MONO remained informative across different disease durations. Meanwhile, NEUT, BASO Ratio, HCT and LYMPH, and albumin were more indicative in the short course (<= 3 years), whereas NLR and chlorine were more indicative in the longer course (> 3 years). This may suggest that, given the overall stability of the model, longitudinal changes in biomarkers should be investigated across different disease courses and age groups.