A NOVEL PROGNOSTIC MODEL INCLUDING LIVER FUNCTION PARAMETERS ACCURATELY PREDICTS 30-DAY MORTALITY IN PATIENTS ADMITTED WITH COVID-19

ABSTRACT

Background and Aims: While the relationship between elevated liver enzymes and COVID- 19 related adverse events is well-established, a liver-dependent prognostic model that predicts the risk of death is helpful to accurately stratify admitted patients. In this study, we use a bootstrapping-enhanced method of regression modeling to predict COVID-19 related deaths in admitted patients. Method: This was a single-center, retrospective study. Univariate and multivariate Cox regression analyses were performed using 30-day mortality as the primary endpoint to establish associated hepatic risk factors. Regression-based prediction models were constructed using a series of modeling iterations with an escalating number of categorical terms. Model performance was evaluated using receiver operating characteristic (ROC) curves. Model accuracy was internally validated using bootstrapping-enhanced iterations. Results: 858 patients admitted to hospital with COVID-19 were included. 78 were deceased by 30 days (9.09%). Cox regression (greater than 20 variables) showed the following core variables to be significant INR (aHR 1.26 95%CI 1.06-1.49), AST (aHR 1.00 95%CI 1.00- 1.00), age (aHR 1.05 95%CI 1.02-1.08), WBC (aHR 1.07 95%CI 1.03-1.11), lung cancer (aHR 3.38 95%CI 1.15-9.90), COPD (aHR 2.26 95%CI 1.21-4.22). Using these core variables and additional categorical terms, the following model iterations were constructed with their respective AUC;model 1 (core only) 0.82 95%CI 0.776-0.82, model 2 (core + demographics) 0.828 95%CI 0.785-0.828, model 3 (prior terms + additional biomarkers) 0.842 95%CI 0.799-0.842, model 4 (prior terms + comorbidities) 0.851 95%CI 0.809-0.851, model 5 (prior terms + life-sustaining therapies) 0.933 95%CI 0.91-0.933, model 6 (prior terms + COVID-19 medications) 0.934 95%CI 0.91-0.934. Model 1 demonstrated the following parameters at 0.91 TPR 0.54 specificity, 0.17 PPV, 0.98 NPV. Bootstrapped iterations showed the following AUC for the respective models model 1 0.82 95%CI 0.765-0.882, model 2 0.828 95%CI 0.764-0.885, model 3 0.842 95%CI 0.779-0.883, model 4 0.851 95%CI 0.808-0.914, model 5 0.933 95%CI 0.901-0.957, model 6 0.934 95%CI 0.901- 0.961. Conclusion: Model 1 displays high prediction performance (AUC >0.8) in both regression-based and bootstrapping-enhanced modeling iterations. Therefore, this model can be adopted for clinical use as a calculator to evaluate the risk of 30-day mortality in patients admitted with COVID-19. (Table Presented)