S260 AI-Based Survival Prediction in Primary Biliary Cirrhosis Using Random Survival Forest and Longitudinal Clinical Features

Introduction: Primary biliary cirrhosis progresses unpredictably and challenges timely intervention. We developed and validated a random survival forest model to predict transplant-free survival and to identify key prognostic markers in a cohort of 418 patients. Methods: We curated a cohort of 418 patients with primary biliary cirrhosis and collected longitudinal data on bilirubin, albumin, copper, prothrombin time, and 13 additional clinical variables. Missing continuous values received mean imputation; missing categorical values received mode imputation. We standardized continuous features to a mean of zero with unit variance and converted categorical variables into binary indicators. Outcomes were defined as death, liver transplantation, or censoring. We randomly allocated 80 percent of the data to a training set and 20 percent to a test set. We then trained a Random Survival Forest with 1,000 decision trees and assessed discrimination via the concordance index (C-index). To identify key prognostic markers, we performed permutation-based feature importance analysis. Results: The Random Survival Forest achieved a training C-index of 0.94 and a testing C-index of 0.88. Permutation analysis highlighted baseline bilirubin and albumin as the most influential predictors, followed by prothrombin time and copper. We generated individualized risk scores and stratified patients into low- (n = 82), moderate- (n = 134), and high-risk (n = 202) groups. Mean risk scores measured 20.1 in the low-risk group, 31.8 in the moderate-risk group, and 93.9 in the high-risk group. Kaplan–Meier curves demonstrated clear separation across all 3 strata, confirming robust prognostic stratification. Conclusion: The model demonstrated high discrimination and successfully stratified patients into distinct risk groups, with bilirubin and albumin emerging as dominant predictors. These results highlight the potential of machine-learning–based risk tools to inform personalized management and warrant external validation for broader clinical application.