A-047 Demographics and clinical features improve prediction of heart failure diagnosis compared to natriuretic peptides alone

Abstract

Acute heart failure (HF) is a leading cause of emergency department (ED) visits in the United States and drives substantial acute care utilization. Despite this burden, diagnostic uncertainty in the ED remains common. Although NT-proBNP is the primary diagnostic biomarker, demographic and clinical characteristics may further improve diagnostic accuracy. We assessed whether incorporating these clinical features enhances HF prediction compared with NT-proBNP alone in a prospective, multicenter ED study. Subjects >18 years old presenting to the ED with symptoms consistent with HF were assessed, with samples tested at admission with the Alere NT-proBNP assay for Alinity i. Excluded subjects were those with renal insufficiency requiring dialysis or eGFR <15 mL/min/1.73 m², trauma, or missing clinical measurements. A total of 1,936 patients (833 HF, 43.0%) were used to train three logistic regression models predicting diagnosis of HF or non-HF. The base model included only NT-proBNP (log2-transformed) as a predictor. The full model included NT-proBNP as well as relevant demographic and clinical features of age, sex, BMI, eGFR category (<= 60 mL/min/1.73 m²), hypertension status, smoking status, and diabetes status as predictors. The reduced model included age, and significant predictors from the full model. Area under the curve (AUC) with 95% confidence interval (CIs) was evaluated for each model, with pairwise comparisons performed using the DeLong test. NT-proBNP showed a significant association with HF diagnosis across all three models. In the fully adjusted model, each doubling of NT-proBNP was associated with an odds ratio of 2.45 (95% CI: 2.27–2.66). Both the full and reduced models outperformed the base model that used NT-proBNP concentration alone to predict diagnosis of HF. In the full model, NT-proBNP, and diabetes status were statistically significant and were included with age and sex as predictors in the reduced model. The base model resulted in an AUC of 0.886 (95% CI: 0.872–0.901), while the full and reduced models achieved AUCs of 0.916 (95% CI: 0.904–0.928) and 0.915 (95% CI: 0.903–0.928), respectively (Figure). The ROC curves of both the full and reduced models differed significantly from that of the base model (p<0.001 in both cases), but not from each other (p=0.25). Accounting for age, and diabetes significantly improves HF prediction beyond NT proBNP alone. Integrating these readily available factors into HF diagnostic algorithms may enhance accuracy over current practice. Such models could aid ED clinicians, enabling faster, more accurate HF diagnosis and more efficient discharge of non HF patients.