A biomarker-based risk score to predict death in patients with atrial fibrillation: the ABC (age, biomarkers, clinical history) death risk score

Abstract

Aims: In atrial fibrillation (AF), mortality remains high despite effective anticoagulation. A model predicting the risk of death in these patients is currently not available. We developed and validated a risk score for death in anticoagulated patients with AF including both clinical information and biomarkers.

Methods and results: The new risk score was developed and internally validated in 14 611 patients with AF randomized to apixaban vs. warfarin for a median of 1.9 years. External validation was performed in 8548 patients with AF randomized to dabigatran vs. warfarin for 2.0 years. Biomarker samples were obtained at study entry. Variables significantly contributing to the prediction of all-cause mortality were assessed by Cox-regression. Each variable obtained a weight proportional to the model coefficients. There were 1047 all-cause deaths in the derivation and 594 in the validation cohort. The most important predictors of death were N-terminal pro B-type natriuretic peptide, troponin-T, growth differentiation factor-15, age, and heart failure, and these were included in the ABC (Age, Biomarkers, Clinical history)-death risk score. The score was well-calibrated and yielded higher c-indices than a model based on all clinical variables in both the derivation (0.74 vs. 0.68) and validation cohorts (0.74 vs. 0.67). The reduction in mortality with apixaban was most pronounced in patients with a high ABC-death score.

Conclusion: A new biomarker-based score for predicting risk of death in anticoagulated AF patients was developed, internally and externally validated, and well-calibrated in two large cohorts. The ABC-death risk score performed well and may contribute to overall risk assessment in AF.

Clinicaltrials.gov identifier: NCT00412984 and NCT00262600.

Keywords: Atrial fibrillation; Biomarkers; Mortality; NOAC; Oral anticoagulation; Risk score.

© The Author 2017. Published by Oxford University Press on behalf of the European Society of Cardiology.

Figures

Figure 1

Relative importance of each variable in the full model. Measured by partial Wald χ2 minus the predictor degrees of freedom. NT-proBNP, N-terminal pro B-type natriuretic peptide; cTnT-hs, cardiac troponin T measured with high-sensitivity assay; GDF-15, growth differentiation factor-15; MI, myocardial infarction; eGFR, estimated glomerular filtration rate; TIA, transient ischaemic attack; AF, atrial fibrillation; df, degrees of freedom.

Figure 2

Nomogram for the final biomarker-based ABC-death risk score. Note that the continuous variables are only represented from the respective 1st to the 99th percentiles. Application of the nomogram is exemplified in Supplementary material online, Figure S9.

Take home figure

Cumulative risk of death by predicted 1-year ABC-death risk group for the derivation (dashed lines, n = 14 611) and the validation (solid lines, n = 8548) data. The vertical bar indicates the 1-year risk.

Figure 4

Decision curve analysis. Net benefit of using a model to predict 1-year event of death as compared with strategies of ‘assume high risk to all’ or ‘assume low risk to all’ for different thresholds. A multivariable model based on all clinical information was used for comparison. The analysis is based on 24 348 patients from the ARISTOTLE and RE-LY trials. ABC-death—Age, Biomarkers (cardiac troponin, NT-proBNP, and GDF-15), Clinical history of heart failure). All clinical information—a model solely consisting of clinical variables (age, gender, smoking, alcohol, prior stroke/TIA, diabetes, hypertension, heart failure, prior myocardial infarction, peripheral arterial disease, vascular disease, AF-type, and prior bleeding). As an example, in a population with approximately 37 deaths per 1000 person-years, for a decision threshold of 5% 1-year risk of death, compared with not using any model the ABC-death model would identify 10 additional true deaths within 1 year per 1000 subjects, without increasing the number of false positive predictions. Not using a model would assume that all subjects have the same risk and is illustrated by the two alternatives of either assuming all are at low risk or that all are at high risk. The corresponding net benefit of using a model with all clinical information is five additional true deaths.

Figure 5

Kaplan–Meier estimated cumulative event rate by randomized treatment (colour) by predicted ABC-death risk classes (panel): 0–1%, 1–2%, and ≥2%.