Accurate illness diagnosis is vital for effective treatment and patient safety. Conventional machine learning models are built on the assumption of balanced medical data to perform cancer diagnoses. However, class imbalance remains a crucial challenge that adversely affects the classifier’s performance and reliability, while the existing ensemble solutions are still prone to noisy data and tend to overlook overlaps near decision boundaries. This paper proposes RE-SMOTEBoost, a double-pruning version of the basic ensemble SMOTEBoost method, designed to overcome these drawbacks. First, the proposed method focuses on generating synthetic samples in overlapping regions to better capture the decision boundary by employing roulette wheel selection. Second, it integrates an entropy filter to reduce noisy data and borderline cases, thereby improving the quality of the generated samples. Third, we propose a double regularization penalty to control the proximity of synthetic samples to the decision boundary and prevent the creation of new overlapping samples. These enhancements enable higher-quality oversampling samples, yielding a more balanced training dataset. Experimental findings demonstrated that the proposed method outperforms state-of-the-art methods, achieving a 3.22 improvement in accuracy and an 88.8 reduction in variance compared to the best-performing methods. Practically, the proposed model provides a robust solution for medical applications, handling data scarcity and imbalance arising from data collection difficulties and privacy constraints.