LightGBM for Liver Disease Detection with Hybrid Hyperparameter Optimization

Abstrak

In response to the growing burden of liver-related disorders, this research develops a supervised learning approach using the Light Gradient Boosting Machine (LightGBM) algorithm to support the early identification of Non-Alcoholic Fatty Liver Disease (NAFLD). The study focuses on constructing and assessing a robust classification model that differentiates individuals with NAFLD from those without the condition based on routinely collected clinical indicators and lifestyle-related characteristics. The dataset, obtained from an open-access NAFLD repository, consists of 1,700 patient records with 10 predictor variables and one binary diagnosis label. The proposed framework employs a stratified shuffle split evaluation scheme with 5-fold and 10-fold cross-validation, using out-of-fold (OOF) probabilities to compute overall performance metrics. The baseline LightGBM model already demonstrated strong performance, achieving 88.94% accuracy, 90.74% precision, 88.99% recall, 89.85% F1-score, and 92.52% AUC under 10-fold cross-validation. To further improve predictive performance, hyperparameter tuning was performed using Optuna and Bayesian Optimization. Among the evaluated approaches, Bayesian-optimized LightGBM achieved the best results, with 93.17% accuracy, 94.49% precision, 92.52% recall, 93.72% F1-score, and 93.28% AUC under 10-fold cross-validation. These findings indicate that systematic hyperparameter optimization can improve the discriminative capability of LightGBM for NAFLD detection and support its potential as a reliable decision-support tool in clinical settings.