Major Depressive Disorder (MDD) is a major global mental health challenge, necessitating timely and precise assessment of its severity to enable appropriate intervention. This study proposes an explainable ensemble learning approach to predict MDD into four severity levels, Mild, Moderate, Severe, and Very Severe, using real-world data from 500 patients collected at a reputed psychiatric department of a medical college in India. The dataset includes 43 features, including socio-demographic, lifestyle, medical, and psychological indicators based on the Hamilton Depression Rating Scale (HAM-D). The methodology involves the implementation of advanced machine learning models, including Random Forest, XGBoost, and ensemble techniques such as soft voting and stacking. To mitigate class imbalance, SMOTE is applied, and model performance is optimized using 5-fold cross-validation with hyperparameter tuning. The stacking ensemble model, with Random Forest as the meta-classifier, achieves outstanding AUC scores: 99% for both Mild and Very Severe, 97% for Moderate, and 96% for Severe. For interpretability, SHAP (Shapley Additive exPlanations) identifies key predictive features. ‘Insomnia - Delayed’ is a prominent indicator for Very Severe and Moderate MDD, ‘Anxiety - Psychological’ is significant for the Severe class, and ‘Feelings of Guilt’ is most influential for Mild cases. This approach provides high predictive accuracy with clinical interpretability, providing a robust tool for personalized MDD severity assessment.