Machine learning to improve the performance of natriuretic peptides across age groups for the diagnosis of acute heart failure

Abstract Background Guidelines recommend N-terminal pro-B-type natriuretic peptide (NT-proBNP) thresholds for the diagnosis of acute heart failure but performance of these thresholds in different age groups is uncertain and could be improved using machine learning methods that incorporate age as a continuous measure. Purpose To evaluate the diagnostic performance of guideline-recommended NT-proBNP thresholds for different age groups and a validated decision-support tool using machine learning for the diagnosis of acute heart failure. Methods NT-proBNP concentrations and the output from the CoDE-HF decision-support tool that uses machine learning to combine NT-proBNP and age as continuous variables with other patient variables were determined in 10,369 patients (median age 73 [interquartile range 59-82]) with suspected acute heart failure pooled from fourteen studies. The diagnostic performance of guideline-recommended NT-proBNP uniform rule-out thresholds (300 pg/mL), and age-stratified rule-in thresholds (450 pg/mL, 900 pg/mL and 1,800 pg/mL for <50, 50-75 and ≥75 years, respectively) and CoDE-HF were evaluated using random effects meta-analysis across patient age groups. Results Overall, 43.9% (4549/10 369) of patients had an adjudicated diagnosis of acute heart failure. The negative predictive value (NPV) of the rule-out threshold of 300 pg/mL was lower in patients with increasing age (NPV 88.2% [confidence interval (CI) 83.5-91.8%] ≥75 years versus 98.9% [97.6-99.5%] <50 years) (Figure 1). Conversely, the positive predictive value (PPV) of age stratified rule-in thresholds was lower in younger patients (PPV 61.8% [55.9-67.4%] <50 years versus 80.5 % [71.1-87.4%] ≥75 years) (Figure 2). CoDE-HF had improved diagnostic performance compared to NT-ProBNP thresholds for all age groups, with the lowest NPV and PPV being 96.4% (93.8-97.9%) and 78.6% (59.8-90.0%), respectively. Conclusion The diagnostic performance of guideline recommended thresholds of NT-proBNP to rule-in and rule-out acute heart failure varies significantly with age. A decision-support tool that incorporates NT-proBNP with age as continuous variables provides a more consistent and accurate approach.Figure 1.Negative predictive valueFigure 2.Positive predictive value