Prediction of long-term survival among patients with cirrhosis using time-varying models

Risk prediction among patients with cirrhosis has historically focused on short-term (ie, 90 days) mortality among patients waitlisted for a transplant. Although several models have been developed to predict intermediate and longer term survivals, they have important limitations, namely, including only baseline laboratory and clinical variables to predict survival over a time horizon of years.

We developed prediction models using time-varying laboratory and clinical data among patients with cirrhosis in the OneFlorida Clinical Research Consortium. We fit extended Cox models and assessed model discrimination and calibration in complete-case analysis and imputation of missing laboratory data.

Among 15,277 patients, 9922 (64.9%) were included in the complete-case analysis. Final models included demographic (age and sex), time-updating laboratory (albumin, alanine transaminase, alkaline phosphatase, bilirubin, platelet, and sodium), and time-updating clinical (ascites, hepatic encephalopathy, spontaneous bacterial peritonitis, and bleeding esophageal varices) variables. Model discrimination was excellent in the complete-case analysis [AUC and concordance-index (C-index) > 0.85] at 1-, 2-, 3-, 4-, and 5-year time points. Model performance was unchanged with the exclusion of race and ethnicity as model predictors. Model discrimination was excellent (C-index >0.8) when imputation was used for patients with 1 or 2 missing laboratory variables.

Using data from a statewide sample of patients with cirrhosis, we developed and internally validated a time-updating model to predict survival with excellent discrimination. Based on its measures of discrimination (AUC and c-index), this model matched or exceeded the performance of other published risk models depending on the time horizon. If externally validated, this risk score could improve the care of patients with cirrhosis by improving counseling on intermediate and longer term outcomes to guide clinical decision-making and advanced care planning.