Revolutionizing CKD Prediction with Bayesian- Optimized LightGBM for Clinical Decision Support
S. Jyothirmaye, Senthilkumar Meyyappan, S. Ilavarasan, M. Muthupandi, G. Vallathan, K. Kasturi
Abstract
Chronic Kidney Disease (CKD) is a progressive condition that requires early detection for effective intervention. Traditional predictive models like Logistic Regression, Decision Trees, and Support Vector Machines suffer from overfitting, poor feature selection, and suboptimal hyperparameter tuning. To address these challenges, a LightGBM-based predictive model optimized with Bayesian Optimization is proposed. It leverages histogram-based learning to efficiently process large datasets, handle missing values and categorical features, and enhance predictive accuracy through optimized feature selection and generalization. Bayesian Optimization further improves model performance by systematically searching for optimal hyperparameters, minimizing bias and variance, and enhancing stability. The proposed LightGBM model achieved significant improvements over traditional methods, with training accuracy increasing from 0.975 to 1 and testing accuracy from 0.975 to 0.99. Additionally, Precision, Recall, and F1-score improved from 0.98 to 0.988, while overall accuracy across CKDGen and UC Irvine datasets increased from 0.98 to 0.987. These enhancements confirm the effectiveness and reliability of the proposed approach for CKD prediction, making it a robust tool for early CKD detection and clinical decision-making.