Feature Extraction and Stacking Model Analysis for Gallstone Disease Prediction based on UCI Dataset

Abstract

Gallstone disease (GD) is a prevalent and multifactorial hepatobiliary disorder with complex etiological mechanisms involving metabolic, genetic, and environmental factors. Accurate early diagnosis is crucial to prevent complications such as cholecystitis, pancreatitis, and bile duct obstruction. However, conventional diagnostic procedures, including ultrasonography and computed tomography, are costly, operator-dependent, and not always feasible for population-level screening. In this study, a comprehensive machine learning framework is proposed to identify and predict gallstone occurrence using the UCI Gallstone Disease Dataset. The methodology integrates advanced data preprocessing, normalization, principal component analysis (PCA), multi-strategy feature selection, and a stacking ensemble learning architecture. After data cleaning and Z-score normalization, PCA was employed to reduce redundancy and extract latent diagnostic features explaining over 95% of the total variance. Three feature selection strategies-mutual information (MI), L1-regularized logistic regression, and random forest (RF) importance-were integrated to select the most discriminative clinical features. A stacking model combining Random Forest (RF), XGBoost (XGB), and Support Vector Machine (SVM) as base learners with Logistic Regression (LR) as the meta-classifier was implemented. The ensemble demonstrated strong generalization performance with cross-validation AUC = 0.8789 ± 0.0283, Accuracy = 0.8120 ± 0.0405, and test AUC = 0.9102, Accuracy = 0.8125, Recall = 0.8438, Specificity = 0.7812, and MCC = 0.6262. These findings indicate that the proposed approach effectively balances sensitivity and specificity, offering a practical computational model for gallstone risk screening and diagnosis.