[Development of a prediction model for acute heart failure based on B-type natriuretic peptide and multiple clinical indicators in acute dyspnea patients with varying renal function levels].

摘要

Objective: To investigate the diagnostic value of B-type natriuretic peptide (BNP) for acute heart failure (AHF) in acute dyspnea patients with different renal function statuses and to develop a diagnostic model for AHF by integrating BNP with multiple clinical variables. Methods: A retrospective analysis was conducted on patients presenting with acute dyspnea as the primary or initial symptom who were admitted to the Emergency Department of Peking University First Hospital between January 2019 and June 2023. Demographic characteristics, clinical features, and laboratory parameters were collected as candidate variables. Renal function was assessed using the estimated glomerular filtration rate (eGFR) and classified into different stages. Least absolute shrinkage and selection operator (LASSO) regression was applied to 34 candidate variables to identify variables associated with AHF diagnosis. Receiver operating characteristic (ROC) curve analysis was performed for the selected variables. The diagnostic performance of BNP across different renal function subgroups was further evaluated using ROC analysis. Subsequently, 5 predictive models, including logistic regression, random forest, Light Gradient Boosting Machine (LightGBM), Extreme Gradient Boosting (XGBoost), and Adaptive Boosting (AdaBoost), were developed based on the selected variables. Model performance was assessed using the area under the curve (AUC) of ROC, calibration curves, and decision curve analysis (DCA). Differences in the AUC across subgroups were compared using the DeLong test. Model interpretability was evaluated using Shapley additive explanations (SHAP). Results: Among the 605 enrolled patients, 236 (39.0%) were diagnosed with AHF and 369 (61.0%) were classified as non-AHF. LASSO regression identified seven variables associated with AHF diagnosis, including history of diuretic use, diabetes mellitus, previous cardiac dysfunction, persistent atrial fibrillation, acute myocardial infarction, eGFR stage, and BNP level. Among the selected variables, BNP demonstrated the highest diagnostic performance for AHF, with an AUC of 0.949 (95%CI: 0.931-0.967), an optimal cutoff value of 488 ng/L, a sensitivity of 80.9%, a specificity of 96.7%, and a Youden index of 0.777. In patients with eGFR≥60 ml·min-1·(1.73 m2)-1 and eGFR<60 ml·min-1·(1.73 m2)-1, the AUC of BNP for diagnosing AHF were 0.944 (95%CI: 0.902-0.987) and 0.938 (95%CI: 0.913-0.962), respectively, with no significant difference between the 2 groups (P=0.793). The logistic regression, random forest, AdaBoost, LightGBM, and XGBoost models achieved AUC of 0.956 (95%CI: 0.937-0.972), 0.944 (95%CI: 0.925-0.962), 0.940 (95%CI: 0.918-0.960), 0.931 (95%CI: 0.909-0.952), and 0.929 (95%CI: 0.905-0.951), respectively. The logistic regression model demonstrated the best overall discrimination, with a sensitivity of 87.7%, a specificity of 91.6%, and a Youden index of 0.793. Hosmer-Lemeshow testing, calibration curves, and DCA further indicated good calibration and clinical utility of the logistic regression model. SHAP analysis identified BNP as the most influential predictor, with the highest mean absolute SHAP value (4.278). Conclusions: BNP exhibited high diagnostic value for AHF regardless of renal function status. The logistic regression model incorporating BNP and routinely available clinical indicators demonstrated excellent diagnostic performance, calibration, and clinical utility, and may serve as a useful tool for assisting the diagnosis of AHF in emergency settings. 目的： 探讨B型利钠肽（BNP）在不同肾功能状态下对急性呼吸困难患者急性心力衰竭（AHF）的诊断价值，并在此基础上结合多项临床指标构建AHF的预测模型。 方法： 回顾性收集并分析2019年1月至2023年6月北京大学第一医院急诊科收治的以急性呼吸困难为主要或首发症状的患者的人口学资料（年龄、性别等）、临床特征（既往心功能不全、糖尿病史、利尿剂使用史等）及实验室指标（BNP、血肌酐等）作为候选诊断变量，根据估算的肾小球滤过率（eGFR）进行肾功能分期。基于34个候选变量采用最小绝对收缩和选择算子（LASSO）回归筛选诊断相关变量，对筛选变量进行单变量受试者工作特征（ROC）曲线分析。进一步采用ROC曲线评价BNP在不同肾功能状态下诊断AHF的效能。在此基础上使用筛选出的变量分别构建二元logistic回归模型、随机森林模型、LightGBM模型、XGBoost模型和AdaBoost模型，并采用ROC曲线下面积（AUC）、校准曲线和决策曲线分析（DCA）评价模型性能。采用DeLong检验比较不同亚组间AUC的差异。采用Shapley加性解释（SHAP）方法评估模型可解释性。 结果： 共纳入605例患者，其中AHF组236例（39.0%），非AHF组369例（61.0%）。使用LASSO回归模型从34个候选变量中筛选出利尿剂使用史、糖尿病史、既往心功能不全、持续性心房颤动、急性心肌梗死、eGFR分期及BNP水平共7个与AHF诊断相关的变量。单变量ROC分析显示，BNP诊断AHF的效能最高，AUC为0.949（95%CI：0.931~0.967），截断值为488 ng/L，敏感度为80.9%，特异度为96.7%，最大约登指数为0.777。在eGFR≥60 ml·min⁻¹·（1.73 m²）⁻¹和eGFR<60 ml·min⁻¹·（1.73 m²）⁻¹患者中，BNP诊断AHF的AUC分别为0.944（95%CI：0.902~0.987）和0.938（95%CI：0.913~0.962），差异无统计学意义（P=0.793）。基于上述7个变量构建的logistic回归、随机森林、AdaBoost、LightGBM和XGBoost模型均具有较好的AHF诊断效能，其AUC分别为0.956（95%CI：0.937~0.972）、0.944（95%CI：0.925~0.962）、0.940（95%CI：0.918~0.960）、0.931（95%CI：0.909~0.952）和0.929（95%CI：0.905~0.951）。其中，logistic回归模型区分度最高，敏感度为87.7%，特异度为91.6%，约登指数为0.793。Hosmer-Lemeshow检验、校准曲线及DCA均提示该模型具有良好的校准能力和临床应用价值。SHAP分析显示，BNP水平为模型贡献度最高的变量，其平均绝对SHAP值为4.278。 结论： BNP在不同肾功能状态下对AHF均具有较高的诊断价值。基于BNP及常规临床指标构建的logistic回归模型具有较好的诊断效能，可作为急诊AHF辅助诊断的参考工具。.