Depression remains a major global health challenge, with diagnosis often hindered by subjectivity, stigma, and severe class imbalance across clinical categories. This paper proposes OSHEDC, an optimized and interpretable hybrid ensemble framework for multi-class depression classification. The objective is to achieve robust predictive performance while maintaining transparency for clinical decision support. OSHEDC integrates Gradient Boosting, Support Vector Machines, and bagged Multi-Layer Perceptrons within a stacking-voting architecture, using Logistic Regression as a meta-classifier. Experiments conducted on a structured survey-based dataset comprising 1,998 samples across twelve depression types demonstrate that OSHEDC outperforms eight baseline models, achieving 99% accuracy, 99% macro-averaged recall and F1-score, and an AUROC of 1.00. To enhance interpretability, LIME is employed to provide instancelevel explanations, revealing clinically meaningful predictors such as Self-Harm, Low Energy, and Education Level. The results indicate that OSHEDC offers a reliable and explainable solution for multi-class depression assessment, with potential applicability in real-world mental health analytics. Future work will explore multimodal data integration and privacy-preserving learning frameworks.