3.1. Population characteristics.
A total of 4550 patients from the Heb-ADHF registry were enrolled into this study, which were randomly divided into the derivation cohort (2,745) and the validation cohort (1,805) at a ratio of 3:2. The incidence of the composite of outcomes in the derivation and validation cohorts were 13.9% (381/2745) and 13.9% (249/1805) respectively. For both cohorts, the age, male, BMI, smoking, drinking, HF duration, number of hospitalizations for heart failure (NH-HF), LoHS, SBP, DBP, heart rate, etiology, heart function, comorbidities, urine protein, laboratory blood tests, chest radiography, electrocardiogram, echocardiography, medical treatment, and risk of outcome events were comparable without significant difference (P > 0.05). The clinical characteristics of derivation and validation cohorts are summarized in Table 1.
3.2. Validation of previous scores in the derivation cohort.
The results showed that discriminations of 3A3B, AHEAD, and OPTIME-CHF score were all poor on the composite of in-hospital all-cause mortality, 30-day all-cause readmission, or 30-day all-cause mortality after discharge in the derivation cohort, with AUROCs of 0.55 (95% CI 0.53-0.57), 0.54 (95% CI 0.53-0.56), and 0.56 (95% CI 0.54-0.57), respectively. Only the specificity of 3A3B score was acceptable (85.3%), but the sensitivities was unsatisfactory (22.0%).
3.3. Logistic regression analysis and the new risk score model.
Logistic regression analysis was performed in the derivation cohort (Table 2). After univariate analysis, there were 27 variables (P < 0.2) entered into the multivariable analysis: sex (male), BMI, LoHS, DBP, anemia, RBC, neutrophil, lymphocyte, platelet, TC, LDL-C, creatinine, BUN, AST, ALT, cTnI/T, BNP (NT-proBNP), QRS fraction, pulmonary congestion (X-ray), LVEF, LAD, ACEI/ARB, β-blocker, Tolvaptan, recombinant human brain natriuretic peptide (rhBNP), calcium channel blockers, and Statin. However, multivariate analysis showed only 8 variables (P < 0.05) [DBP, lymphocyte, creatinine, BUN, BNP (NT-proBNP), QRS fraction of electrocardiogram, ACEI/ARB, and rhBNP] maintained an independent correlation with the composite endpoint, which were included into the development of the final model.
We calculated the relative importance of each independent predictor by the value of (χ2﹣df) (Figure 1). Then, the score was assigned based on the (χ2﹣df) value of each independent predictor (Table 2). The new scoring system was as follows:
1 × (DBP < 80 mmHg) + 2 × (lymphocyte > 1.11 × 109/L) + 1 × (creatinine > 80 μmol/L) + 2 × (BUN > 21 mg/dL) + 1 × [BNP 500 to < 1500 pg/mL (NT-proBNP 2500 to < 7500 pg/mL)] or 3 × [BNP ≥ 1500 (NT-proBNP ≥ 7500) pg/mL] + 3 × (QRS fraction of electrocardiogram < 55%) + 4 × (ACEI/ARB not used) + 1 × (rhBNP used).
3.3. Discrimination, calibration, and comparison of the new risk score.
The discrimination ability of the new PRS model was suitable, whose AUROC for the derivation cohorts was 0.67 (95%CI 0.64-0.70) (Figure 2A), and it had good calibration (the Hosmer-Lemeshow test χ2=3.366, P=0.186). The discrimination was validated by the AUROC for the validation cohorts [0.65 (95% CI 0.61-0.69)] (Figure 2B), and the calibration was also good (the Hosmer-Lemeshow test χ2=9.751, P=0.283). In addition, the AUROCs of the new PRS in both the derivation and validation cohorts were the highest compared with the 3A3B score, AHEAD score, and OPTIME-CHF score system (Figure 2), those differences were statistically significant (P < 0.001; Table 3).
3.4. The risk stratification of the new risk score model in the entire cohort.
The new PRS was then used to classify patients into 3 groups: 0-4 points, low-risk group; 5-8 points, medium-risk group; ≥ 9 points, high-risk group. The incidence of in-hospital all-cause mortality in low-, medium-, and high-risk group were 1.86% (14/754), 4.07% (74/1820), and 5.57% (110/1976), respectively (Figure 3A). The incidence of 30-day readmission in low-, medium-, and high-risk group were 3.18% (24/754), 6.43% (117/1820), and 11.03% (218/1976), respectively (Figure 3B). The incidence of 30-day all-cause mortality after discharge were 0.53% (4/754), 1.04% (19/1820), and 2.53% (50/1976), respectively (Figure 3C) and the composite of outcome events were 5.57% (42/754), 11.54% (210/1820), and 19.13% (378/1976), respectively (Figure 3D). Mantel-haenszel test showed a linear trend between the risk stratification and the incidence of outcomes events (χ2=38.14, P < 0.001). Pearson correlation test indicated that the incidence of outcomes events increased with increase of the risk stratification (R=0.480, P < 0.001).