A Hybrid Feature Selection and Hyperparameter-Tuned XGBoost Approach for Liver Disease Prediction

Liver disease is becoming a major global health issue, highlighting the importance of early detection for timely treatment and better patient outcomes. This study presents a hybrid machine learning framework that combines advanced preprocessing techniques, a dual feature selection strategy— Mutual Information (MI) followed by Recursive Feature Elimination (RFE)—and a genetically optimized XGBoost classifier to predict liver disease using the Indian Liver Patient Dataset (ILPD). The process starts with comprehensive preprocessing steps, including categorical encoding, imputation of missing values, class balancing through SMOTE, outlier management via Winsorization, and Box-Cox transformation to address skewness. A two-step feature selection method is employed: initially, MI identifies the top features, which are then refined using RFE with an XGBoost estimator to select the most informative subset. The optimized XGBoost model achieved a test accuracy of 85.03%, a recall of 90.67%, and a ROC-AUC score of 91.09%, demonstrating strong predictive capabilities. The confusion matrix reveals a low number of false negatives (7), while the ROC curve visually confirms the model's discriminative power. Additionally, SHAP-based explanations enhance transparency by identifying the most influential features affecting the predictions. This explainable, high-performing hybrid framework offers a practical and interpretable solution for predicting liver disease, aiding clinicians in making early and accurate diagnoses. The approach demonstrates strong accuracy and robustness, paving the way for future integration of explainable AI tools in clinical decision-making systems.