- Patient characteristics
A total of 943 HBV-related cirrhosis and acute gastrointestinal bleeding were included, of whom 629 and 314 patients were randomly enrolled into the derivation and validation cohorts, respectively. Table 1 compares demographic, clinical, and laboratory characteristics between the two cohorts of patients. The two cohorts were not entirely matched: in the derivation cohort, the percentage of PRBC transfusion was lower and endoscopic treatment was higher than those among the patients in the validation cohort (Table 2).
- Lipid profiles correlate with liver disease severity
Lipid profiles include triglyceride, cholesterol, HDL-C and low-density lipoprotein cholesterol (LDL-C). In the derivation and validation cohorts, the association of individual lipid fractions with prognostic models (Child-Pugh, MELD, and ALBI) was assessed by Pearson correlation, as shown in Table 3. HDL-C was found to have strongest correlation with liver function scores.
- Derivation cohort: predictors of 6-week mortality
On ROC analysis, HDL-C showed excellent diagnostic accuracy for 6-week mortality, with an AUROC of 0.847 (95% CI 0.789-0.905). The best cut-off value of HDL-C was 0.54mmol/L, with a sensitivity of 85.1% and specificity of 74.2%.
We performed univariate and multivariate analysis to assess the independent predictors of 6-week mortality. Variables included in the univariate analysis are shown in Table 4. The variables significantly associated with 6-week mortality in the univariate analysis were as follows: age, systolic blood pressure, ascites, hepatic encephalopathy, HCC, heart rate, PRBC transfusion, endoscopic treatment, hemoglobin (HGB), total leukocyte count (WBC), alanine aminotransferase (ALT), serum sodium (Na), total bilirubin (TBIL), albumin (ALB), QTc interval prolongation, cholesterol, HDL-C, LDL-C, serum creatinine (Scr), international normalized ratio (INR), fibrinogen (FIB) (Table 4). Multivariate logistic regression analyses showed that TBIL, HDL-C, Na, and HGB were independently associated with 6-week death (Table 4). Based on the multivariate regression coefficients, we calculated a new prognostic model for HBV-related cirrhotic patients with acute GIB named N-CGIB according to the following formula: 0.005 × TBIL (mmol/L) – 0.112 × Na (mmol/L) – 2.757 × HDL (mmol/L) – 0.017 × HGB (g/L) + 16.121. The N-CGIB model showed an excellent discrimination for 6-week death prediction, with an AUROC of 0.895 (95% CI 0.853-0.936). Hosmer and Lemeshow analysis confirmed goodness-of-fit for the model (P=0.117).
The developed model showed an excellent predictive accuracy, with AUROCs significantly better than that of Child-Pugh, MELD, ALBI, D’Amico score, Augustin model, AIMS65 score and Glasgow-Blatchford score (Table 5, Fig. 1A).
- Mortality in external validation cohort
HDL-C was an excellent predictor of 6-week mortality (AUROC 0.827; 95% CI 0.752-0.901) in the validation cohort.
Hosmer and Lemeshow analysis confirmed the goodness-of-fit of the N-CGIB model (P=0.066). The diagnostic accuracy of the N-CGIB model (AUROC 0.893; 95% CI 0.850-0.937) was similar to D’Amico model (AUROC 0.836; 95% CI 0.765-0.907), significantly better than that of the other models, including Child–Pugh (AUROC 0.757; 95% CI 0.673-0.842), MELD (AUROC 0.797; 95% CI 0.712-0.881), ALBI (AUROC 0.808; 95% CI 0.744-0.872), Augustin model (AUROC 0.766; 95% CI 0.667-0.866), AIMS65 score (AUROC 0.773; 95% CI 0.697-0.849), and Glasgow-Blatchford (AUROC 0.753; 95% CI 0.669-0.837) (Table 5, Fig 1B).
- Discrimination of N-CGIB for variceal bleeding and non variceal bleeding
We also analyzed the prognostic performance of N-CGIB in the subset with variceal hemorrhage considering the whole group of enrolled patients (536 of 953; 56.2%). As seen, N-CGIB score (AUROC 0.914, 95% CI 0.878-0.951) showed an excellent predictive accuracy for AVB. The discriminatory performance of the N-CGIB model (AUROC 0.841; 95% CI 0.783-0.896) was lower for NVB as compared to that for AVB.
- Survival
In the derivation cohort, 66 of the 629 (10.5%) patients died during study observation, with early death rate (16 of 66, 24.2%) occurring within the first five days. In the test set, 34 of the 314 (10.8%) patients died during study observation, with early death rate (7 of 34, 20.5%) occurring within the first five days.
Mortality increased with increasing N-CGIB score. The final selected cutoff value accurately discriminated patients in 3 subgroups with distinct prognosis: a low-risk group (N-CGIB score below -3), whose in-hospital mortality was 1.7%; an intermediate-risk group (N-CGIB score range from -3 to -1), with a 12.3% mortality; and a high-risk group (N-CGIB score more than -1) associated with a 56.9% or greater predicted risk. Comparable results were found in the validation set with a significant decrease in 6-week probability of survival: 1.1%, 17.9%, and 54.2%, respectively. (Fig. 2A.2B)
On Kaplan-Meier analysis, baseline HDL-C values ≤ 0.54mmol/L were associated with markedly lower 6-week survival in the derivation cohorts (Fig. 3A), and validation cohorts (Fig. 3B).