SVM-ECOC Driven Multi-Class Survival Prediction in Primary Biliary Cirrhosis Through Comparative Evaluation of Ensemble and Kernel-Based Machine Learning Approaches

Abstract

Primary biliary cirrhosis is a chronic autoimmune liver disease that progresses through fibrotic stages and ultimately threatens patient survival. Accurate prediction of patient outcomes, including death, censoring, and liver transplantation, remains a clinically meaningful yet statistically challenging task due to severe class imbalance, high rates of missing clinical data, and overlapping feature distributions among outcome groups. In this study, we present a systematic comparative analysis of six machine learning classifiers applied to a well-characterized cohort of 418 patients from the Mayo Clinic. The classifiers examined include Random Forest, AdaBoostM2, RUSBoost, Bagged Trees, Support Vector Machine with Error-Correcting Output Codes, and Subspace K-Nearest Neighbors. All models were evaluated under a rigorous framework incorporating stratified five-fold cross-validation, leakage-free preprocessing, and multi-metric assessment across accuracy, macro-averaged F1-score, precision, and recall. Among the six methods, the SVM-ECOC classifier achieved the highest test accuracy of 0.8072 and the best overall average ranking of 1.75 across all four metrics, while RUSBoost attained the highest macro F1-score of 0.5810 and was the only method capable of detecting the minority transplant class. These findings highlight the fundamental tension between overall classification accuracy and equitable per-class sensitivity in imbalanced clinical datasets, and they offer practical guidance for the design of prognostic models in hepatology.