Ensemble learning has emerged as a powerful approach for disease diagnosis, combining multiple classifiers to enhance predictive accuracy and robustness. Nevertheless, the challenge lies in selecting an optimal ensemble configuration while balancing accuracy and diversity. This study introduces a Bagging Approach with Teaching-Learning-Based Optimization (BA-TLBO) algorithm for ensemble optimization in disease diagnosis. To strike a balance between accuracy and diversity, a novel fitness function is proposed. This function incorporates ensemble mean accuracy and mean diversity, utilizing the Hamming distance as a measure of diversity. Additionally, dynamic weight updating is suggested to optimize weights over iterations in the BA-TLBO optimization process, thereby balancing exploration and exploitation. The use of a dynamic bag size over iterations aims to balance bias and variance, thereby enhancing generalization. The BA-TLBO explores different classifier combinations iteratively by selecting and replacing bags within the ensemble. This process aims to achieve high accuracy while also maintaining diversity. The effectiveness of the proposed approach is tested on four benchmark disease diagnosis datasets using multiple classifiers, including Logistic Regression (LR), K-Nearest Neighbors (KNN), Decision Trees (DT), and Support Vector Machines (SVM). The model’s performance is compared using diversity metrics, including Entropy, Bhattacharya distance, and Q statistics. Results indicate the superiority of the proposed model over alternative approaches. Furthermore, the robustness of the proposed model is compared with other meta-heuristic optimization algorithms, such as Artificial Bee Colony (ABC), Ant Colony Optimization (ACO), Firefly Optimization (FO), and Particle Swarm Optimization (PSO). The evidence suggests that the proposed model performs better in the majority of cases, particularly in 5-bag and 10-bag configurations. The proposed approach is evaluated using both 5-bag and 10-bag configurations, considering both worst-case and best-case bag optimization strategies. Experimental results demonstrate that the BA-TLBO-based model outperforms both state-of-the-art (SOTA) ensemble and non-ensemble models.