Abstract 4357888: Evaluating the Temporal and Sociodemographic Generalizability of the Emergency Heart Failure Mortality Risk Grade

Background: The Emergency Heart Failure Mortality Risk Grade (EHMRG) models are a series of clinically validated models to predict (7-day and 30-day) mortality in acute heart failure (AHF) patients. Although EHMRG models were validated in 2015, it is unknown whether EHMRG maintained temporal validity, and if they maintain accuracy in different sociodemographic strata. Objectives: To assess the temporal validity of EHMRG models using recently collected data and determine if EHMRG models perform differently across sociodemographic subgroups Methods: We selected a sample of 7537 patients presenting to 10 emergency departments with an AHF diagnosis between 2017-2019 in Ontario Canada, from the Comparison of Outcomes and Access to Care for Heart Failure Trial (COACH). We linked our cohort to national databases to obtain census-derived sociodemographic factors. We calculated the EHMRG-7 (7-day mortality), and EHMRG30-ST (30-day mortality) scores for each patient, and assessed the performance of those models using discrimination metrics and calibration plots comparing the observed and predicted probabilities of death across the deciles of risk. We then compared the discriminatory performance across several demographic subgroups. Results: In our independent temporal validation cohort (median age 80 [IQR, 70-87] years, 48% females), EHMRG-7 had a high discrimination c-statistic = 0.795 [95%CI, 0.764 – 0.824] and demonstrated good calibration with minor overestimation of the 7-day risk of death in the highest decile. EHMRG30-ST had a good discrimination, with a c-statistic = 0.777 [95%CI, 0.759 – 0.7969], but overestimated the risk of 30-day death across the upper five deciles. The discrimination of both EHMRG models was robust across subgroups of sex, marital status, income quintiles, and census derived factors, as demonstrated by comparable c-statistics in these subgroups. However, the c-statistic was lower for patients residing in retirement or long-term care homes. (Figure 1) Conclusions: EHMRG models retained good discrimination comparable to previous validations. Risk was overestimated at 30 days, which would lead to more conservative recommendations tending towards hospital admission. EHMRG models were robust over the majority of sociodemographic characteristics.