Comprehensive Summary
Doudesis and their colleagues explored whether machine learning can improve the diagnostic use of natriuretic peptides in patients presenting with symptoms that could represent acute heart failure. Current guidelines recommend fixed thresholds for BNP and MR proANP, but these values often fluctuate based on age, kidney function, body mass index, and more. To understand this variation, the investigators pooled individual patient data from fourteen international studies, including 8493 BNP cases and 3899 MR proANP cases. BNP at 100 pg/mL yielded a negative predictive value (NPV) of 93.6%and a positive predictive value (PPV) of 68.8%. MR proANP at 120 pmol/L had an NPV of 95.6 percent and a PPV of 64.8%. Accuracy dropped in patients with obesity, atrial fibrillation, COPD, and prior heart failure. The authors developed CoDE HF, a machine learning tool that incorporates BNP or MR proANP as continuous values along with routine clinical features. In patients without prior heart failure, the BNP model reached an AUROC of 0.914, and the MR proANP model reached an AUROC of 0.929.
Outcomes and Implications
This study shows that fixed natriuretic peptide cutoffs may not meet the needs of real-world emergency care. CoDE HF provides a more individualized probability that supports safer rule-out decisions and helps prevent missed cases in high-risk groups. Therefore, clinicians can apply this tool to rapidly identify patients who can be discharged and those who require faster imaging, cardiology consultation, and early heart failure therapy. Broader adoption could reduce unnecessary admissions, improve diagnostic equity across diverse patient groups, and support earlier treatment that improves outcomes.