Hybridized light gradient boosting and whale optimization algorithm for diabetes detection

Dada, Emmanuel Gbenga and Birma, Aishatu Ibrahim and Gora, Abdulkarim Abbas and Okunade, Oluwasogo Adekunle and Hassan, Abubakar (2025) Hybridized light gradient boosting and whale optimization algorithm for diabetes detection. World Journal of Advanced Engineering Technology and Sciences, 15 (2). pp. 1966-1981. ISSN 2582-8266

Due to their remarkable precision and effectiveness, gradient-boosted tree models have become the go-to choice for machine learning-driven diabetes detection; however, the key to unlocking their full potential lies significantly in the careful tuning of hyperparameters. To automatically optimize LGBM's hyperparameters for improved diabetes screening, we present a hybrid framework - Light Gradient Boosting (LGBM) bundled with the Whale Optimization Algorithm (LGBM+WOA). Inspired by nature, the Whale Optimization Algorithm (WOA) models the bubble-net feeding behaviour of humpback whales, therefore offering a compromise between exploration and exploitation in search areas. We evaluated model performance under imbalanced class situations using stratified 10-fold cross-validation using the Diabetes Dataset from patients in Borno hospital. Rising above baseline Gradient Boosting (80%), Support Vector Machine (74%), Random Forest (86%), and LGBM (88%), the suggested LGBM+WOA model achieved an overall detection accuracy of 90%. While diabetes recall increased to 0.86, so lowering false negatives is important; class-specific metrics for the non-diabetic cohort obtained a precision of 0.93, recall of 0.91, and F1-score of 0.92 - gains of 1-2 percentage points over standard LGBM. Faster convergence and better generalization follow from WOA-driven hyperparameter tuning, refining important LGBM parameters more effectively than grid or random search. The easier training and testing process of the hybrid model is a helpful tool for quickly assessing diabetes risk and allows for immediate use in clinical decision support systems. Combining LGBM's gradient-boosting efficiency with WOA's robust global optimization, the LGBM+WOA framework provides a new benchmark for machine-learning-based diabetes detection, enabling more general uses of metaheuristic-tuned ensembles in medical diagnostics.