Comparison of Mortality Prediction Scores in Intermediate-Care Patients with Liver Cirrhosis at a German University Transplant Centre: A Prospective Study

Abstract

<b><i>Background and Aims:</i></b> Mortality prediction models help to extract and relate patient data upon admission to intensive or intermediate care units (ImCUs). Considering technical and economic healthcare developments, re-evaluations of score performances are required to warrant their validity. This study validates and compares established scoring systems in cirrhotic ImCU patients. <b><i>Methods:</i></b> Acute Physiology and Chronic Health Evaluation (APACHE) II, Simplified Acute Physiology Score (SAPS) 2 and 3, Sepsis Organ Failure Assessment (SOFA), Mortality Probability Model at ICU admission (MPMo) II and III, Model for End stage Liver Disease (MELD), CLIF-Consortium Acute-on-Chronic Liver Failure (CLIF-C ACLF), CLIF-Consortium Acute Decompensation (CLIF-C AD), and Intermediate Care Unit Severity Score (ImCUSS) were calculated in patients with cirrhosis (<i>n</i> = 98) at ImCU admission. Discrimination performances were evaluated by area under the receiver operating characteristic curves (AUROCs), calibration performances with calibration belt plots, and their corresponding <i>p</i> values. <b><i>Results:</i></b> Overall, SAPS 3 and CLIF-C ACLF have shown the best 90-day mortality prediction outcomes with AUROCs of 0.825 and 0.783 along with calibration belt <i>p</i> values of 0.128 and 0.061, respectively. In a subgroup analysis of patients with acute-on-chronic liver failure (ACLF), expanded SAPS 2, SOFA, and SAPS 3 reached the best AUROCs, i.e., 0.760, 0.750, and 0.714, but none of the tested scores reached an acceptable calibration. <b><i>Conclusion:</i></b> Ninety-day mortality risk prediction of the SAPS 3 and CLIF-C ACLF was accurate in our cohort of patients with liver cirrhosis admitted to ImCUs. A particular challenge remains that is the mortality prediction in patients with ACLF requiring ImCU-level care; here, further developments are needed to generate scores with acceptable predictive performances.