Liver Disease Prediction Using Machine Learning: A Leakage Free Evaluation of SVM, Logistic Regression, Random Forest and XGBoost on the Indian Liver Patient Dataset

Liver disease prediction using machine learning has been studied widely on the Indian Liver Patient Dataset (ILPD), yet many reported accuracies come from a single train and test split, from resampling carried out before the data are divided, or from duplicated patient records that appear in both training and test data. This study presents a leakage free evaluation of five classifiers for liver disease detection from routine liver function tests: support vector machine (SVM), logistic regression, random forest, XGBoost and k nearest neighbours (kNN). After removing 13 duplicate records, 570 patients were analysed. Imputation, log transformation, scaling and minority oversampling were fitted inside each training fold, and performance was estimated by nested cross validation with three repeats of stratified fivefold outer validation and a fivefold inner grid search. Logistic regression reached the highest mean ROC AUC of 0.754 (95% CI 0.717 to 0.790) and the SVM reached 0.741 (0.683 to 0.800). No pairwise difference was significant under the corrected resampled t test with Holm adjustment. A preliminary single split experiment with the same SVM had given an AUC of 0.806, which shows how weak evaluation inflates published results. Handling class imbalance mattered more than the choice of classifier, since the SVM fell to an AUC of 0.690 without it. Every model detected disease less well in women than in men (SVM AUC 0.629 against 0.785). SHAP values and permutation importance ranked alanine aminotransferase, bilirubin, aspartate aminotransferase and alkaline phosphatase as the strongest predictors. On a second liver dataset for hepatitis C related disease, every model exceeded an AUC of 0.97, and the final SVM pipeline scores one patient in 1.6 ms. These findings give a realistic baseline for machine learning liver disease screening and show that sex stratified validation is needed before clinical use.