Abstract

Early diagnosis and correct identification of liver disease is paramount in improving patient outcomes and helping with prompt clinical interventions. Despite high reported predictive accuracy on the structured liver datasets on machine learning models, questions have been raised on transparency of validation, overfitting risk, interpretability and computational feasibility. This paper is a statistically rigorous, interpretable ensemble-learning model assessed by a large Kaggle dataset of 30691 patient records. The pipeline suggested combines preprocessing, two filtering steps of features, imbalance reduction based on Edited Nearest Neighbors and optimization of the hyperparameters via 5-fold cross-validation. Four ensemble models have been comparatively evaluated on Accuracy, F1-score, Matthews Correlation Coefficient, ROC-AUC, PR-AUC and 95% confidence intervals. XGBoost was the best performing model with the highest testing accuracy of 99.57%, cross-validation mean accuracy of 99.85% with minimal standard deviation and the highest MCC of 0.9893. Besides predictive power, XGBoost took 0.58 seconds to train and 0.17 seconds to infer, which is better in terms of computational efficiency. SHAP and LIME explainability analysis established that the model decisions depend on clinically defined liver biomarkers. The findings show that statistical rigor, interpretability, and efficiency are effective concepts in combination to predict liver disease in a reliable and clinically meaningful manner.