Machine learning to aid in the diagnosis of acute heart failure in the emergency department

Abstract

Abstract Background B-type natriuretic peptide (BNP) and mid-regional pro-atrial natriuretic peptide (MRproANP) testing are recommended to aid in the diagnosis of acute heart failure. However, the application of these biomarkers for optimal diagnostic performance is uncertain. Methods We performed a systematic review and harmonised individual patient-level data to evaluate the diagnostic performance of BNP and MRproANP for the diagnosis of acute heart failure using random-effects meta-analysis. We subsequently developed and externally validated a decision-support tool called CoDE-HF for both BNP and MRproANP that combines the natriuretic peptide concentrations with clinical variables using machine learning to report the probability of acute heart failure for an individual patient. Results Fourteen studies from 12 countries provided individual patient-level data in 8,493 patients for BNP and 3,847 patients for MRproANP, in whom, 48.3% (4,105/8,493) and 41.3% (1,611/3899) had an adjudicated diagnosis of acute heart failure, respectively. The negative and positive predictive values of guideline-recommended thresholds for BNP (100 pg/mL) and MR-proANP (120 pg/mL) were 93.6% (95% confidence interval 88.4–96.6%) and 68.8% (62.9–74.2%), and 95.6% (92.2–97.6%) and 64.8% (56.3–72.5%), respectively. However, we observed significant heterogeneity in the diagnostic performance across important patient subgroups (Figure 1). In the external validation cohort, CoDE-HF was well calibrated with excellent discrimination in those without prior acute heart failure for both BNP and MRproANP (area under the curve of 0.946 [0.933–0.958] and 0.943 [0.921–0.964], and Brier scores of 0.105 and 0.073, respectively). CoDE-HF performed consistently across all subgroups for both BNP and MRproANP, and identified 30% and 65.7% at low-probability (negative predictive value of 99.1% [98.8–99.3%] and 99.1% [98.8–99.4%]), and 30% and 17.3% at high-probability (positive predictive value of 91.3% [90.7–91.9%] and 70.0% [68.5–71.4%]) in those without prior heart failure, respectively (Figure 2). Conclusion In an international collaborative analysis, we observed that guideline-recommended thresholds for BNP and MRproANP to diagnose acute heart failure varied significantly across patient subgroups. A decision-support tool using machine learning to combine natriuretic peptides as a continuous measure and other clinical variables provides a more accurate and individualised approach. Funding Acknowledgement Type of funding sources: Other. Main funding source(s): Medical Research Council and British Heart Foundation Figure 1. NPV of BNP threshold (100 pg/mL)Figure 2. NPV of the CoDE-HF rule-out score