S2841 AI-Based Risk Stratification and Time-to-Transplant Prediction in Cirrhosis Using Random Survival Forest Modeling

Introduction: Cirrhosis patients often follow an unpredictable clinical course, and timely identification of those at greatest risk for liver transplantation remains an unmet need. We developed a random survival forest combining routine laboratory values with 3 engineered predictors age-adjusted enzyme ratios, treatment-response flags, and a composite severity score to predict transplant-free survival and identify key prognostic markers. Methods: We assembled a de-identified cohort of cirrhosis patients from a tertiary‐care database and extracted routinely collected variables, including age, bilirubin, serum glutamic-oxaloacetic transaminase, albumin, comorbidity indices, and treatment history. We engineered 3 novel features—age-adjusted liver enzyme ratios, binary treatment-response flags, and a composite severity score—and handled missing data via mean imputation for continuous variables and mode imputation for categorical variables. We randomly allocated 80% of the records to a training set and reserved 20% for independent testing. We then implemented a random survival forest with 100 trees using the sksurv library, tuning split-rule and node-size hyperparameters through 5-fold cross-validation. We assessed discrimination by the concordance index and visualized risk stratification with Kaplan–Meier survival curves. We applied permutation-based importance analysis to rank predictors of time to transplantation. Results: The model yielded a concordance index of 0.94 on the training cohort and 0.82 on the test cohort. Permutation importance pinpointed age, bilirubin, and SGOT as the top 3 predictors of transplant risk. Stratification into low-, medium-, and high-risk groups captured 29%, 37%, and 34% of the cohort, respectively. Median transplant-free survival was 2,450 days in the low-risk group, 1,520 days in the medium-risk group, and 890 days in the high-risk group. Kaplan–Meier curves demonstrated clear separation among all 3 risk strata (log-rank P < 0.001), confirming robust prognostic discrimination and supporting the model’s potential for guiding timely transplant assessments. Conclusion: Our model demonstrated high discriminative performance, stratifying patients into low-, medium-, and high-risk groups with significantly different median transplant-free survival times. Incorporating novel engineered features alongside traditional laboratory measures shows promise for personalized transplant referral strategies; external validation will be essential to confirm clinical utility.