Prospective validation of a machine learning algorithm for the diagnosis of acute heart failure

前瞻性验证用于诊断急性心力衰竭的机器学习算法

阅读:1

Abstract

BACKGROUND: Diagnosing acute heart failure in the Emergency Department remains challenging. CoDE-HF (Collaboration for the Diagnosis and Evaluation of Heart Failure) is a previously developed machine learning algorithm that predicts an individualised probability of acute heart failure using NT-proBNP as a continuous variable in combination with routinely available clinical variables.(1) PURPOSE: To externally validate the CoDE-HF algorithm in a prospective cohort of patients presenting with acute breathlessness. METHODS: Patients presenting to the Emergency Department with acute breathlessness in whom acute heart failure was suspected were enrolled in this study. CoDE-HF integrates NT-proBNP with age, estimated glomerular filtration rate, haemoglobin, body mass index, heart rate, systolic blood pressure, peripheral oedema, chronic obstructive pulmonary disease and ischaemic heart disease. Diagnosis was adjudicated by two independent cardiologists, with disagreements resolved by a third. Performance was assessed using the area under the receiver operating characteristic curve (AUC), Brier score, and diagnostic accuracy at pre-specified thresholds. RESULTS: Overall, 1,030 patients (mean age 73±14, 49% female) were included in this study, of whom, 29% (294/1,030) had a prior diagnosis of heart failure. Acute heart failure was adjudicated in 374 of 1,030 patients (36%). In patients without prior heart failure, the NT-proBNP rule-out threshold of 300 pg/mL identified 32% as low probability, with sensitivity 95.9% (91.9–97.9) and negative predictive value (NPV) 96.6% (93.3–98.3). The age-specific rule-in thresholds identified 47% as high probability, with specificity 65.4% (61.3–69.3) and positive predictive value (PPV) 45.8% (40.6–51.1). In the same group, CoDE-HF had an AUC of 0.839 (95% CI 0.809–0.869) and Brier score of 0.152. A rule-out threshold of 5.7 identified 30% as low probability, with sensitivity 98.4% (95.3–99.5) and NPV 98.6% (95.7–99.5). A rule-in threshold of 69.2 identified 19% as high probability, with specificity 90.1% (87.2–92.3) and PPV 61.7% (53.4–69.3). In patients with prior heart failure, the NT-proBNP age-specific rule-in thresholds identified 75% as high probability, with specificity 33.6% (25.5–42.8) and PPV 65.9% (59.4–71.9). The CoDE-HF rule-in threshold of 85.0 identified 36% as high probability, with specificity 80.5% (72.2–86.8) and PPV 80.2% (71.7–86.6). CONCLUSIONS: In this prospective cohort, CoDE-HF provided improved, individualised diagnostic accuracy, especially in patients without prior heart failure. It offers a consistent, data-driven tool to support clinical decision-making.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。