Background: Early prediction of liver disease remains challenging in routine clinical diagnostics due to the multifactorial nature of hepatic dysfunction and the limited discriminative capacity of conventional laboratory-only assessments. Objective: This study aims to develop and rigorously evaluate a robust machine learning framework for binary liver disease classification, emphasizing predictive stability, diagnostic balance (sensitivity–specificity), and statistical reproducibility across repeated experiments. Methodology: A structured dataset of 1,700 records with 11 features representing demographic, behavioral, genetic, and clinical determinants was used to train and compare five supervised models: CatBoost, AdaBoost, Random Forest, Support Vector Machine (SVM), and Decision Tree. Performance was assessed under two stratified train–test partitions (60:40 and 70:30). Each experimental configuration was independently repeated 20 times, and mean ± standard deviation was reported for Accuracy, Precision, Recall, F1-score, Specificity, and ROC-AUC to quantify reliability and cross-run robustness. Confusion matrices and ROC curves were used for complementary diagnostic interpretation. Results: Ensemble learners consistently outperformed non-ensemble counterparts across both partitions. CatBoost exhibited the strongest and most stable performance, achieving approximately 0.88 Accuracy and 0.95 ROC-AUC, alongside the lowest false negative rate (∼0.11). The narrow standard deviations across the 20-run protocol indicate high reproducibility and reduced sensitivity to random sampling effects, supporting the generalization strength of the proposed evaluation design. Conclusion & Contribution: The study provides a statistically grounded, reproducible ML evaluation framework for early liver disease prediction and demonstrates that optimized ensemble learning, particularly CatBoost, can enhance diagnostic accuracy while reducing clinically critical misclassification. These findings support the feasibility of AI-assisted screening pipelines and establish a methodological foundation for future translation into non-invasive decision-support systems in preventive hepatology.
Ali et al. (Fri,) studied this question.