Predictive analytics, which utilise advanced practices, tools, and methodologies, are essential for early detection and analysis of various medical diseases. The integration of artificial intelligence (AI) into healthcare has notably improved the performance of disease classification models. On the other hand, current research indicates that individual classification models may yield lower accuracy under specific conditions. This work addresses the significant issues of poor generalization as well as inadequate accuracy in complex medical datasets, where independent machine learning algorithms frequently fail. Hence, this research proposes a tri-level ensemble approach that combines bagging, boosting, and stacking to classify various medical diseases. The ensemble approach involves Gradient Boosting (GB), Extreme Gradient Boosting (XGB), Light Gradient Boosting Machine (LGBM), and CatBoost. Leveraging the strengths of different ensemble methods enhances model performance and achieves higher accuracy. The data are initially obtained from the repository and preprocessed using the StandardScaler. Then, the significant features from the processed data are selected by using the Generalised Opposition Learning-based Osprey Optimization Algorithm (GOL-OOA) approach. Eventually, the tri-level ensemble approach is employed to classify the medical data, and the proposed approach achieves an average accuracy of 97.00%, 96.45%, 99.96%, 98.83% and 94.74% on Pima, Yeast, Leukemia, Breast Cancer and Statlog datasets, respectively, compared to the iterative Neighbourhood Component Analysis (iNCA) and Cooperative binary-real GA-based Extreme Learning Machine (CGA-ELM) and other ensemble approaches.