Prediction of Depression Severity by Applying Machine Learning to Blood Biomarkers

Depression is a complex mental disorder that negatively impacts an individual’s general well-being and daily routine. The diagnosis of depression levels generally relies on the patient’s self-reports and clinical assessments. Analyses relying on biological data are more dependable and can provide early diagnosis to protect the patient from severe consequences. This study investigates the potential of using blood biomarkers to predict depression severity (mild-severe) by applying various Machine Learning (ML) techniques. Five different machine learning methods Logistic Regression (LR), k-Nearest Neighbor (KNN), Decision Tree (DT), Random Forest (RF), and Multilayer Perceptron (MLP) were applied to a dataset including 7,326 samples and 107 features provided by Adana Dr. Ekrem Tok Mental Health Hospital. Class imbalance in the training set was eliminated by the Synthetic Minority Oversampling Technique (SMOTE) method, and the performance of the models was measured by the Receiver Operator Characteristic (ROC) curve and Area Under the Curve (AUC). Random Forest achieved the best performance with an AUC of 0.82 after SMOTE dataset. Feature importance analysis before and after SMOTE indicated that TSH, HCV Ab, and Chlorine were among the most important indicators, ranking in the top five in both cases. The results demonstrate that blood biomarkers can be used as reliable predictors for estimating depression severity.