Explainable Clinical Decision Support for Metabolic Index Prediction in Gout Patients Using GA-Optimized Ensemble Learning Models

Background/Objectives: Metabolic indices such as HOMA-IR, METS-IR and TyG play a critical role in determining cardiometabolic risk and insulin resistance in clinical practice. This study aims to simultaneously predict these three indices using ensemble learning models, based on demographic, clinical, and laboratory data from patients with gout. Methods: Demographic, retrospective clinical, and laboratory data from 411 patients diagnosed with gout were analyzed, and a logarithmic transformation was applied to address skewness in the target variable. A genetic algorithm was used to identify relevant features for predicting metabolic indices, and four scenarios were designed. Five ensemble learning models were optimized using fine-tuning with Optuna. The models’ decisions were clinically validated using SHAP analysis. Results: Across all scenarios, the CatBoost model combined with the GA + Log-Transformed framework achieved the highest prediction accuracy and the lowest error rates. Notably, when predicting the TyG index, it simulated the underlying formula with near-perfect accuracy. SHAP analysis confirmed that CatBoost successfully mapped clinical pathophysiology by prioritizing insulin levels for HOMA-IR, BMI for METS-IR, and triglyceride variance for the TyG index. Conclusions: The proposed GA + Log-Transformed CatBoost workflow provides a highly accurate, non-invasive and interpretable clinical decision support system. This study demonstrates that the ensemble architecture effectively captures complex pathophysiological patterns aligned with medical knowledge rather than merely memorizing statistical noise.