Machine learning to optimize use of natriuretic peptides in the diagnosis of acute heart failure.

Summary

Across 14 studies (8,493 BNP; 3,899 MR-proANP), guideline thresholds showed variable performance by subgroup. The machine learning-based CoDE-HF tool, combining peptide levels with clinical variables, achieved excellent discrimination (AUC ~0.91–0.93), identified large low- and high-probability subsets with NPV ~98.5% and PPV ~75–79%, and maintained calibration across subgroups.

Key Findings

• Guideline thresholds for BNP (100 pg/mL) and MR-proANP (120 pmol/L) yielded NPV 93.6%/95.6% and PPV 68.8%/64.8%, with significant subgroup heterogeneity.
• CoDE-HF achieved AUC 0.914 (BNP) and 0.929 (MR-proANP), with good calibration (Brier 0.110/0.094).
• CoDE-HF identified 30–48% as low probability (NPV 98.5%) and 28–30% as high probability (PPV 75–79%), consistent across subgroups.

Clinical Implications

CoDE-HF can guide rule-in/rule-out decisions by integrating natriuretic peptides with clinical context, identifying low- and high-probability patients with high NPV/PPV. Implementation studies are needed to assess workflow integration, equity, and impact on outcomes.

Limitations

• Prospective clinical impact and workflow integration not yet tested in randomized implementation trials
• Model performance may depend on availability/quality of clinical variables and may require EHR integration

Future Directions

Prospective, multicenter implementation studies to assess clinical utility, health equity, and outcome impact; evaluation in patients with prior heart failure and diverse settings; regulatory and EHR integration.