Comparing the ability of GRACE 2.0 and GRACE 3 scores for predicting in-hospital mortality in patients with and without diabetes

Abstract Background Sex-specific refinement of GRACE 2.0 was published as GRACE 3.0 (G3). The relative ability of G3 in predicting in-hospital mortality in diabetic patients has not been explored. Purpose Explore the ability of G3 in determining in-hospital mortality in patients with and without diabetes. Methods Retrospective cohort analysis of data on NSTEMI and UA patients admitted in two hospitals between 2010 and 2019 entered into the Myocardial Infarction National Audit Project database. All baseline characteristics and G2, G3 scores were recorded. In hospital death was recorded from hospital records. The discriminative power of the G2 score was determined by comparing the area under the receiver operating characteristic curve (c-statistics). Accuracy of calibration was evaluated by plotting the mean predicted vs observed mortality in each decile of predicted risk, Hosmer-Lemeshow (HL) goodness-off it test and the Nagelkerke R2 (NR2). Results G3 was calculated in 2768 patients (age 69±14yrs, 62.3% male, 877(31.7%) with diabetes. Median risk for the cohort was 1.3%, (95% CI 1.2 to 1.4). G3 risk was higher in patients with diabetes (median 2.0 IQR 0.6 to 5.5 vs 1.0 IQR 0.3 to 3.7, p<0.0001). In-hospital mortality occurred in 68 (2.5%) patients with no difference between the diabetes groups (3.1% and 2.2% died in the DM and non-DM groups respectively). In the whole cohort, G3 reclassified 139(7.5%) patients in low-intermediate to high risk-category and 159(17.4%) patients in high to low-intermediate category. 10.1% and 6.5% were upgraded and 16.1% and 18.1% were downgraded in DM and non-DM groups. Overall discriminatory power of G3 (AUC 0.87, 95% CI 0.86 to 0.88, p<0.0001) was excellent but no different from G2 (AUC 0.86 95% CI 0.85 to 0.88, p<0.0001) (δAUC 0.009, p=0.461). In patients with diabetes, G3 yielded higher AUC (AUC 0.85, 95% CI 0.83 to 0.87, p<0.0001) than G2 (AUC 0.82, 95% CI 0.79 to 0.84, p<0.00101) but this was not statistically significant (δAUC 0.031, p=0.142). In patients without DM, the difference was smaller (δAUC 0.002, p=0.906). Neither G2 (δAUC 0.061, p=0.185) nor G3 (δAUC 0.028, p=0.473) was better in discrimination in DM compared to non-DM groups (Fig1A,B). In the diabetic group, calibration plot for the G3 score showed suboptimal calibration with consistent overestimation of the risk across all risk deciles (Figure1C), with the HL test showing significant differences between the observed and predicted values (p <0.0001) and the NR2 was low 0.09. The GRACE 2 score showed better calibration (Fig1D) (HL: p=0.09, NR2 0.11). In patients without diabetes, calibration plot for the G3 (HL: p= 0.17; NR2 0.33) (Fig1E) and G2 (HL: p=0.15, NR2 0.25) (Fig1F) shows similar pattern of overestimation of the risk less so for latter. Conclusion G3 and G2 shows excellent discrimination for in-hospital mortality in both diabetic and non-diabetic patients. G2 seems to show better calibration than G3 for diabetic patients.