A predictive analytics approach with Bayesian-optimized gentle boosting ensemble models for diabetes diagnosis

Abstract

Effective disease management necessitates the accurate and timely prediction of lung cancer and diabetes. Machine learning (ML) based models have garnered attention in the realm of predictive healthcare, with ensemble methods, in particular, bolstering algorithms to improve classification performance. Nevertheless, enhancing boosting algorithms to achieve superior predictive accuracy continues to be a difficult task. This study proposes a Bayesian-Optimized GentleBoost Ensemble (BOGBEnsemble) to improve classification performance for diabetes prediction (DiP) and lung cancer prediction (LCP). Two Kaggle datasets—a diabetes dataset from multiple healthcare providers and a Survey Lung Cancer dataset from existent medical records—are utilized. Data preprocessing involves outlier removal, min–max normalization, class balancing, and Pearson correlation-based feature selection. The GentleBoost classifier is optimized using Bayesian hyperparameter tuning, focusing on learning rate and the number of weak learners, and is validated using 10-fold cross-validation. BOGBEnsemble is evaluated in comparison to leading models, such as Random Forest (RF), Adaptive Boosting (AdaBoost), Logistic Boosting (LogitBoost), Random Undersampling Boosting (RUSBoost), conventional GentleBoost, and Multi-Layer Perceptron (MLP) architectures. The DiP-BOGBEnsemble achieves a 99.26% accuracy, 98.94% precision, 99.60% recall, 99.26% F1-score, 99.46% F2-score, 98.51% MCC, 98.51 Kappa, 0.0041 FOR, and 22,606.75 DOR. The LC-BOGBEnsemble achieves a 96.51% accuracy, 97.83% precision, 94.76% recall, 96.28% F1-score, 95.36% F2-score, MCC of 93.03%, Kappa of 92.99, FOR of 0.0462, and DOR of 932.15. This study highlights the potential of BOGBEnsemble as a clinically viable tool for early disease detection and decision support, paving the way for more reliable and personalized healthcare strategies.