OPTIMASI MODEL LIGHTGBM UNTUK PREDIKSI DEMAM BERDARAH DENGUE MENGGUNAKAN OPTUNA PADA RSUD SYEKH YUSUF GOWA
TL;DR
The proposed Optuna-LightGBM framework provides high diagnostic stability and serves as an effective Clinical Decision Support System (CDSS) for medical practitioners in acute dengue stratification.
Abstract
Dengue Hemorrhagic Fever (DHF) remains a significant infectious disease threat in tropical countries, including Indonesia. Early identification of severe clinical conditions based on medical records is crucial to prevent life-threatening plasma leakage complications. This study aims to construct an optimal classification model for predicting DHF clinical risk levels (Low Risk without Warning Signs vs. High Risk with Warning Signs) using the Light Gradient Boosting Machine (LightGBM) optimized with the Optuna framework. The secondary dataset was collected from RSUD Syekh Yusuf Gowa, comprising 267 clean clinical records partitioned with stratified sampling (80% training and 20% testing). Optuna with Tree-structured Parzen Estimator (TPE) and 5-Fold Stratified Cross-Validation was deployed across 50 trials to tune LightGBM hyperparameters by maximizing the Macro F1-score. The experimental results demonstrated that Optuna optimization significantly elevated the model performance compared to the baseline LightGBM: Accuracy increased from 90.74% to 92.59%, Precision Macro from 84.24% to 86.67%, Recall Macro from 81.11% to 86.67%, and F1-Score Macro from 82.55% to 86.67%. On the test set, the final model achieved 95.56% precision and recall for High-Risk cases and 77.78% for Low-Risk cases. Feature importance analysis revealed that platelet count (trombosit), leukocyte count (leukosit), body temperature, fever duration, and age were the most decisive clinical predictors. The proposed Optuna-LightGBM framework provides high diagnostic stability and serves as an effective Clinical Decision Support System (CDSS) for medical practitioners in acute dengue stratification.