利用膳食抗氧化剂识别心血管疾病人群合并慢性肾病：基于机器学习与SHAP分析

Identification of Coexisting Chronic Kidney Disease among Individuals with Cardiovascular Disease Using Dietary Antioxidants: A Machine Learning and SHAP Analysis

摘要: 背景与目的：慢性肾病(CKD)是心血管疾病(CVD)人群中的常见共病，并与不良预后密切相关。然而，CVD人群中合并CKD的关键识别特征尚未充分明确，膳食抗氧化剂在疾病识别中的作用也缺乏系统评估。本研究旨在采用可解释机器学习方法构建CVD人群CKD识别模型，并分析膳食抗氧化剂及其他相关特征对模型输出的贡献。方法：本研究基于美国国家健康与营养调查(NHANES) 2007~2008年、2009~2010年及2017~2018年数据，纳入年龄 ≥ 20岁的CVD参与者，排除妊娠、膳食资料不可靠、能量摄入异常及关键变量缺失者。CKD判定标准为估算肾小球滤过率 < 60 mL/min/1.73 m²或尿白蛋白肌酐比 ≥ 30 mg/g。采用逻辑回归、随机森林、极端随机树、梯度提升树、XGBoost和LightGBM构建CKD识别模型，并以AUC-ROC、AUC-PR等指标评价模型性能。进一步采用SHAP方法解释膳食抗氧化剂及其他关键特征对模型输出的贡献。结果：共纳入946名CVD参与者，其中单纯CVD组560人，CVD合并CKD组386人。XGBoost模型的区分能力最佳，AUC-ROC为0.852，AUC-PR为0.726；在默认阈值0.50下，其准确率、敏感度和特异度分别为0.742、0.781和0.723。SHAP分析显示，年龄、中高强度体力活动、糖尿病、高血压及体质指数等基础特征对模型输出贡献较大。膳食抗氧化剂中，芹菜素、矢车菊素、槲皮素、维生素E、镁、染料木素和柚皮素等的SHAP重要性相对较高；部分连续特征与模型输出之间呈现非线性关系。结论：在CVD人群中，除年龄、体力活动及代谢性疾病等基础特征外，部分膳食抗氧化剂也可为同期CKD状态识别提供补充信息。基于XGBoost的可解释机器学习模型具有较好的区分能力，但本研究为横断面分析，结果仅用于识别参与者当前是否合并CKD，不能用于预测未来CKD发生。未来仍需通过纵向研究和外部样本进一步验证。

Abstract: Background and Objective: Chronic kidney disease (CKD) is a common comorbidity among individuals with cardiovascular disease (CVD) and is closely associated with adverse outcomes. However, the key features for identifying coexisting CKD in the CVD population have not been fully clarified, and the role of dietary antioxidants in disease identification has not been systematically evaluated. This study aimed to develop an interpretable machine learning model for identifying CKD among individuals with CVD and to assess the contributions of dietary antioxidants and other relevant features to model output. Methods: This study used data from the 2007~2008, 2009~2010, and 2017~2018 cycles of the National Health and Nutrition Examination Survey (NHANES). Participants aged ≥ 20 years with CVD were included. Those who were pregnant or had unreliable dietary data, implausible energy intake, or missing key variables were excluded. CKD was defined as an estimated glomerular filtration rate < 60 mL/min/1.73 m² or a urinary albumin-to-creatinine ratio ≥ 30 mg/g. Logistic regression, random forest, extremely randomized trees, gradient boosting, XGBoost, and LightGBM were used to develop models for identifying CKD. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC-ROC), the area under the precision–recall curve (AUC-PR), and other performance metrics. SHapley Additive exPlanations (SHAP) was further used to interpret the contributions of dietary antioxidants and other key features to model output. Results: A total of 946 participants with CVD were included, of whom 560 had CVD alone and 386 had coexisting CKD. XGBoost showed the best discrimination, with an AUC-ROC of 0.852 and an AUC-PR of 0.726. At the default threshold of 0.50, its accuracy, sensitivity, and specificity were 0.742, 0.781, and 0.723, respectively. SHAP analysis showed that age, moderate-to-vigorous physical activity, diabetes, hypertension, and body mass index made relatively large contributions to model output. Among dietary antioxidants, apigenin, cyanidin, quercetin, vitamin E, magnesium, genistein, and naringenin showed relatively high SHAP importance. Several continuous features also showed nonlinear relationships with model output. Conclusion: Among individuals with CVD, several dietary antioxidants may provide supplementary information for identifying concurrent CKD, in addition to basic features such as age, physical activity, and metabolic conditions. The interpretable XGBoost model showed good discriminatory ability. However, because this was a cross-sectional analysis, the model is intended only to identify current CKD status and cannot be used to predict future CKD onset. Further validation in longitudinal studies and external samples is needed.