第一部分 心力衰竭患者血压、代谢及睡眠指标的预后价值分析

研究目的：血压与代谢在心力衰竭的发生和发展过程中发挥着关键作用，一方面，心力衰竭患者存在“血压悖论”现象；另一方面，胰岛素抵抗与心血管疾病密切相关。此外，睡眠障碍在心力衰竭患者中也较为常见。目前，关于血压、代谢和睡眠指标综合应用于心力衰竭患者预后评估的研究相对有限。本研究旨在分析收缩压（systolic blood pressure，SBP）、舒张压（diastolic blood pressure，DBP）、反映胰岛素抵抗的甘油三酯-葡萄糖（triglyceride-glucose，TyG）指数和TyG-体重指数（body mass index，BMI）及反映睡眠的呼吸暂停低通气指数（apnea hypopnea index，AHI）对心力衰竭患者预后的影响，并探索血压、代谢及睡眠指标联合应用在心力衰竭预后评估中的潜在价值。

研究方法；本研究为一项回顾性队列研究，纳入2006年7月至2023年12月于中国医学科学院阜外医院心力衰竭重症监护病房住院，血压、代谢及睡眠指标相关资料完整，且定期规律随访的心力衰竭患者。研究的主要终点为全因死亡，次要终点包括心血管死亡和心血管不良事件。采用Spearman相关性分析法评估血压、代谢和睡眠指标之间及其与临床特征之间的关系。采用单因素及多因素Cox回归分析，并结合Kaplan-Meier生存曲线，评估SBP、DBP、TyG、TyG-BMI和AHI与患者终点事件之间的相关性。此外，通过一致性评分（concordance index，C-index）、准确率、召回率、精确率、F1分数以及受试者工作特征曲线下面积（area under the curve，AUC）等评估方法，基于基础心力衰竭预后模型，该模型包括年龄、性别、估算肾小球滤过率（estimated glomerular filtration rate，eGFR）、左室射血分数（left ventricular ejection fraction，LVEF）、纽约心脏协会（New York Heart Association，NYHA）分级、血管紧张素转化酶抑制剂（angiotensin-converting enzyme inhibitor，ACEI）或血管紧张素受体拮抗剂（angiotensin receptor blocker，ARB）或血管紧张素受体脑啡肽酶抑制剂（angiotensin receptor-neprilysin inhibitor，ARNI）治疗、β受体阻滞剂治疗，进一步联合应用血压、代谢和睡眠指标，评估其对心力衰竭患者预后评估的增益效应。

研究结果：本研究共纳入2196例心力衰竭患者，中位随访时间为762天，中位年龄57岁，期间548例患者发生主要终点事件。Spearman相关性分析结果显示，SBP、TyG、TyG-BMI和AHI指标之间呈现显著正相关。多因素Cox回归中，校正年龄、性别、eGFR、LVEF、NYHA分级、ACEI/ARB/ARNI及β受体阻滞剂治疗后，SBP升高与心力衰竭患者的全因死亡（风险比[hazard ratio，HR] 0.985，95%置信区间[confidential interval，CI] 0.978-0.992，P<0.001）、心血管死亡（HR 0.983，95%CI 0.974-0.991，P<0.001）及心血管不良事件（HR 0.993，95%CI 0.987-0.999，P=0.016）风险降低相关。TyG升高与全因死亡（HR 0.643，95%CI 0.442-0.845，P<0.001）及心血管死亡（HR 0.701，95%CI 0.467-0.936，P=0.003）风险降低相关。TyG-BMI升高也与全因死亡（HR 0993，95%CI 0.990-0.996，P<0.001）和心血管死亡（HR 0994，95%CI 0.990-0.997，P<0.001）风险下降相关。DBP和AHI与全因死亡、心血管死亡及心血管不良事件的关系未达统计学意义。在亚组分析中，高龄、男性、合并冠心病、较高NYHA分级、较高水平的N末端B型利钠肽前体和高敏肌钙蛋白I均与不良预后密切相关。进一步分析表明，在基础心力衰竭预后模型的基础上，逐步添加血压、代谢及睡眠指标能够显著提高模型的C-index和AUC，且改进效果具有统计学意义，有效提升了模型对于心力衰竭患者预后的预测效能。

研究结论：SBP、胰岛素抵抗标志物TyG和TyG-BMI升高与心力衰竭患者的全因死亡和心血管死亡风险降低相关，其中SBP升高还与心血管事件风险的下降密切相关。DBP和AHI与全因死亡、心血管死亡及心血管不良事件之间的关系不显著。在基础心力衰竭预后评估模型的基础上，联合血压、代谢和睡眠指标能够提升模型对全因死亡、心血管死亡和心血管不良事件的预测效能。

第二部分 基于可解释性机器学习的心力衰竭患者短期预后模型构建

研究目的：心力衰竭作为心血管疾病的终末期状态，不仅严重影响患者的生活质量，还给家庭和社会带来了沉重的经济和护理负担。此外，心力衰竭患者面临着高死亡率和高再住院率的问题，其不良预后已成为临床面临的重要挑战。传统的风险评估方法存在一定局限性，如依赖单一生物标志物、未能充分考虑多维数据的复杂交互关系等。本研究基于可解释性机器学习模型，一方面探索机器学习算法在特征筛选中的应用价值，另一方面整合多种临床变量，构建用于心力衰竭患者的风险预测工具。本研究在优化预测性能的同时，关注模型的可解释性，从而提升模型的可用性和可信度，并基于验证后的模型尝试开发网页计算器，提供辅助决策支持。

研究方法：本研究为一项回顾性队列研究，训练集和测试集数据来源于重症监护医学信息数据库IV（Medical Information Mart for Intensive Care IV，MIMIC-IV），外验证数据集来源于eICU协作研究数据库（eICU Collaborative Research Database，eICU-CRD），纳入确诊为心力衰竭且首次住院、住院时长在24小时至28天之间的患者。研究主要终点为在院全因死亡。采用Python和R语言进行数据分析和可解释性机器学习模型构建。在特征筛选阶段，采用最小绝对值收敛和选择算子（least absolute shrinkage and selection operator，LASSO）回归算法筛选关键变量。随后构建15种机器学习算法模型，包括线性判别分析（linear discriminant analysis，LDA），二次判别分析（quadratic discriminant analysis，QDA），逻辑回归，岭回归分类器（ridge classifier），轻量级梯度提升机（light gradient boosting machine，LightGBM），极致梯度提升（eXtreme Gradient Boosting，XGBoost)，梯度提升，决策树，随机森林，极致随机森林（extremely randomized trees，Extra Trees），自适应增强（adaptive boosting，AdaBoost），朴素贝叶斯，虚拟分类器，K近邻算法（K-nearest neighbors，KNN）和线性核支持向量机（linear kernel support vector machine，SVM）。模型评估采用多项指标，包括准确率、召回率、精确率、F1分数、受试者工作特征曲线下面积（area under the curve，AUC）、科恩卡帕分数（Cohen's kappa score）和马修斯相关系数（Matthews correlation coefficient）。在所有模型中，选取性能最佳的模型，并基于沙普利加和解释（SHapley Additive exPlanations，SHAP）方法进行可解释性分析，以揭示关键变量及其对模型决策的影响。另一方面，基于单因素和多因素逻辑回归，构建进一步精简变量的列线图（nomogram）模型。最佳模型和列线图均通过eICU-CRD进行外部验证，并在二者的基础上，分别开发网页计算器和动态列线图计算器。

研究结果：本研究共纳入2343例心力衰竭患者，其中535发生终点事件。研究人群的中位年龄为70岁，男性患者约占60%（1413例），中位住院天数为6天。与终点事件组相比，未发生事件的患者年龄较小，住院时间更短。研究共选取75个临床特征，涵盖基本信息、生命体征、实验室检查、血流动力学指标和血氧状态等多维度变量。在特征筛选过程中，采用LASSO回归算法确定正则化参数α=0.020，最终筛选出44个特征用于模型构建。在15个机器学习模型的比较中，LDA模型综合性能最佳。具体而言，LDA模型在训练集的准确率为0.8354，测试集的准确率为0.8563，并在AUC、召回率、精确率、F1分数、科恩卡帕分数和马修斯相关系数等多个评估指标上均表现良好。LDA模型在训练集和测试集上的一致性指数（concordance index，C-index）分别为0.7972和0.8125。在外部验证集的进一步测试表明，LDA模型在准确率、召回率、精确率和F1分数方面均接近0.9，AUC达到0.79。此外，在训练集中进行单因素逻辑回归分析，并将具有统计学意义的变量纳入多因素逻辑回归模型，进一步构建列线图。最终列线图模型纳入14个关键特征，其在训练集、测试集及外部验证集的AUC分别为0.852、0.855和0.770，同时校准曲线显示出良好的预测一致性。研究基于LDA模型和列线图部署了网页计算器。

研究结论：机器学习模型在预测住院心力衰竭患者在院24小时至28天的全因死亡风险方面表现良好，显示了预后评估的准确性和实用性。SHAP方法增强了模型的可解释性。研究部署了LDA模型及动态列线图网页计算器，提供了便捷的且高效在线风险评估工具。

第三部分 中老年人群代谢指标与血压对心血管风险影响的递进式分层分析

研究目的：心血管疾病（cardiovascular disease，CVD）是全球面临的重要健康挑战，尤其对中老年人群构成严重威胁。在众多潜在的危险因素中，甘油三酯-葡萄糖（triglyceride-glucose，TyG）指数等代谢标志物作为反映胰岛素抵抗的重要指标，可能在评估CVD风险中具有重要作用。本研究旨在探讨TyG、TyG-体重指数（body mass index，BMI）及其累积指标等代谢指标与CVD风险之间的关系，并进一步评估不同代谢暴露水平下血压对CVD风险的影响，为CVD风险评估和管理策略提供科学依据。

研究方法：本研究为一项回顾性队列研究，基于中国健康与养老追踪调查（China Health and Retirement Longitudinal Study，CHARLS）数据集，针对基线时未患有CVD的参与者进行分析。CHARLS每两年进行一次随访调查，截至目前共进行了五期随访。本研究基于基线（第一期）调查收集患者基线信息，结合第三期和第五期的随访数据判断新发CVD情况。研究使用Cox比例风险回归模型，构建不同校正水平的模型，纳入年龄、性别、户口信息、教育程度、婚姻状况、吸烟史、高血压和肾脏疾病等协变量，分析TyG、TyG-BMI、累积TyG及累积TyG-BMI等代谢指标与CVD风险的关系。同时，应用限制性立方样条回归（restricted cubic spline regression，RCS）分析方法评估代谢指标与CVD风险之间的非线性关系。通过递进式分层分析，根据累积TyG和累积TyG-BMI水平将患者划分为低代谢暴露组和高代谢暴露组，探讨不同代谢负荷下血压对CVD风险的影响。

研究结果：本研究共纳入4800例参与者，其中女性2651例，占55.23%，男性2149例，占44.77%，中位年龄为57岁。与基线时相比，第三期随访时参与者的BMI、TyG和TyG-BMI均显著升高。CVD组和无新发CVD组在血压、代谢指标及合并症等方面均存在显著差异。Cox比例风险回归分析显示，TyG、累积TyG、TyG-BMI及累积TyG-BMI与长期CVD风险显著相关，且随访时间越长，影响越强。在4年随访期，TyG与CVD风险的关系未达到统计学显著性（HR=1.019，95%CI 0.891-1.165，P=0.785），而在9年期时，TyG对风险的影响增强，风险增加13.4%，且具有统计学意义（HR=1.134，95%CI 1.026-1.253，P=0.014）。累积TyG指数呈现出类似的趋势，4年随访期（HR=1.014，95%CI 0.962-1.068，P=0.607）的结果并不显著，而9年时风险增加达5.9%（HR=1.059，95%CI 1.019-1.102，P=0.004）。对于TyG-BMI，每增加10单位，在4年和9年随访期中会导致风险增加3-4%。累积TyG-BMI对CVD风险的影响较为温和，4年和9年随访期中，每增加10单位会导致CVD风险增加1-2%。将代谢指数按照四分位数分组后，分析发现TyG、累积TyG、TyG-BMI和累积TyG-BMI处于较高四分位数的个体，其CVD风险升高。相比于Q1、Q2和Q3等指标处在较低范围的参与者，具有较高指标的参与者面临的长期CVD风险更高。进一步的RCS回归分析中，TyG和累积TyG与CVD风险呈线性正相关，即代谢水平越高，CVD风险越大。而TyG-BMI和累积TyG-BMI则展现出倒U型关系，在中等水平时对CVD风险的影响最为显著。代谢暴露分层分析发现，低代谢负荷时，血压升高对CVD风险的影响更强。而在高代谢负荷下，血压与CVD风险的关联性相对较弱，且需要更长时间方可显现。此外，无论在低代谢还是高代谢负荷状态下，舒张压对CVD风险的影响均大于收缩压。

研究结论：TyG和累积TyG与中老年人群长期CVD风险呈线性正相关，而TyG-BMI和累积TyG-BMI则表现为倒U型关系，中等水平对风险的影响最为显著。随着随访时间的延长，代谢指标在CVD风险评估中的作用更加突出。血压在低代谢负荷下对CVD风险的影响更强，而在高代谢负荷下则相对减弱，且舒张压对CVD风险的影响大于收缩压。

Part 1: Prognostic value of blood pressure, metabolic, and sleep indicators in heart failure

Objective: Blood pressure and metabolism play critical roles in heart failure (HF). On the one hand, the “blood pressure paradox” has been observed in HF patients; on the other hand, insulin resistance is closely associated with cardiovascular diseases. Additionally, sleep disorders are prevalent among HF patients. However, studies that integrate blood pressure, metabolic, and sleep indicators for prognostic assessment in HF remain limited. This study aims to analyze the impact of systolic blood pressure (SBP), diastolic blood pressure (DBP), triglyceride-glucose (TyG) index, TyG-body mass index (TyG-BMI), and apnea-hypopnea index (AHI) on HF prognosis and explore the potential value of integrating these traits in prognosis prediction.

Methods: This retrospective cohort study included 2196 HF patients hospitalized in the heart failure care unit (HFCU), Fuwai Hospital between July 2006 and December 2023. Patients with complete records of blood pressure, metabolic, and sleep indicators and who had regular follow-ups were enrolled. The primary endpoint was all-cause mortality, while secondary endpoints included cardiovascular mortality and major adverse cardiovascular events (MACE). Spearman’s correlation analysis was adopted to evaluate the relationships between blood pressure, metabolic, and sleep indicators and their associations with clinical characteristics. The associations between SBP, DBP, TyG, TyG-BMI, and AHI and patient outcomes were assessed using univariate and multivariate Cox regression analyses combined with the Kaplan-Meier method. Furthermore, the added prognostic value of incorporating blood pressure, metabolic, and sleep indicators into a baseline HF prognostic model-comprising age, sex, estimated glomerular filtration rate (eGFR), left ventricular ejection fraction (LVEF), New York Heart Association (NYHA) classification, and treatments with angiotensin-converting enzyme inhibitor (ACEI) or angiotensin receptor blocker (ARB) or angiotensin receptor neprilysin inhibitors (ARNI), and β-blockers-was evaluated using the concordance index (C-index), accuracy, recall, precision, F1 score, and area under the receiver operating characteristic curve (AUC).

Results: There were 2,196 patients included in the study, with a median follow-up of 762 days and median age of 57. During follow-up, 548 patients experienced the primary endpoint. Spearman’s correlation analysis showed significant positive correlations among SBP, TyG, TyG-BMI, and AHI. In multivariate Cox regression analysis, after adjusting for age, sex, eGFR, LVEF, NYHA classification, treatments ACEI/ARB/ARNI, and β-blocker therapy, the increase in SBP was associated with a reduced risk of all-cause mortality (hazard ratio [HR] 0.985, 95% confidence interval [CI] 0.978-0.992, P<0.001), cardiovascular mortality (HR 0.983, 95% CI 0.974-0.991, P<0.001), and MACE (HR 0.993, 95% CI 0.987-0.999, P=0.016). Higher TyG levels were associated with lower all-cause mortality (HR 0.643, 95% CI 0.442-0.845, P<0.001) and cardiovascular mortality (HR 0.701, 95% CI 0.467-0.936, P=0.003) risks. Similarly, higher TyG-BMI was correlated with lower all-cause mortality (HR 0.993, 95% CI 0.990-0.996, P<0.001) and cardiovascular mortality (HR 0.994, 95% CI 0.990-0.997, P<0.001) risks. However, DBP and AHI were not significantly associated with all-cause mortality, cardiovascular mortality, or MACE. In subgroup analyses, older age, male sex, concomitant coronary artery disease, higher NYHA classification, elevated N-terminal pro-B-type natriuretic peptide, and high-sensitivity cardiac troponin I were all significantly associated with worse prognosis. Further analysis demonstrated that the stepwise addition of blood pressure, metabolic, and sleep indicators to the baseline model led to a significant increase in the C-index and AUC, demonstrating a statistically significant improvement in the performance for HF prognosis evaluation.

Conclusion: Higher SBP, TyG, and TyG-BMI levels were significantly associated with a lower risk of all-cause mortality and cardiovascular mortality in HF patients. Additionally, increased SBP was closely related to a reduced risk of MACE. However, DBP and AHI showed no significant associations with all-cause mortality, cardiovascular mortality, or MACE. Integrating blood pressure, metabolic, and sleep indicators into HF prognostic models can enhance the performance of HF prognosis evaluation.

Part 2: Construction of explainable machine learning models for short-term prognosis evaluation in heart failure

Objective: Heart failure (HF) is the end stage of cardiovascular disease, significantly reducing patients' quality of life while imposing burdens on families and society. Moreover, HF is associated with high mortality and rehospitalization rates, making adverse prognosis a critical challenge. Traditional risk assessment methods have limitations, such as relying on single biomarkers or the inability to account for complex interactions among multidimensional clinical data. This study aims to develop a prediction model for HF based on an explainable machine-learning method. Specifically, this study explores the value of machine learning algorithms in feature selection while integrating multiple clinical variables. The study focused on both performance optimization and model interpretability to enhance reliability. To assist in decision-making, two web calculators were established for applications based on validated models.

Methods: This retrospective cohort study utilized data from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database and eICU Collaborative Research Database (eICU-CRD), involving HF patients hospitalized for the first time, with the length of stay between 24 hours and 28 days. The primary outcome was in-hospital all-cause mortality. Data analysis and model construction were performed using Python and R. Feature selection was conducted using the least absolute shrinkage and selection operator (LASSO) regression algorithm, with the optimal regularization parameter set at α=0.020, ultimately selecting 44 key features for model development. And 15 machine learning models were constructed, including linear discriminant analysis (LDA), quadratic discriminant analysis (QDA), logistic regression, ridge classifier, light gradient boosting machine (LightGBM), eXtreme gradient boosting (XGBoost), gradient boosting, decision tree, random forest, extremely randomized trees (Extra Trees), adaptive boosting (AdaBoost), naive Bayes, dummy classifier, K-nearest neighbors (KNN), and linear kernel support vector machine (SVM). Model performance was evaluated using accuracy, recall, precision, F1 score, area under the receiver operating characteristic curve (AUC), Cohen’s kappa score, and Matthews correlation coefficient (MCC). The best-performing model was selected for further interpretability analysis using the SHapley Additive exPlanations (SHAP) method to identify key predictors and their influence on the model. Additionally, a simplified nomogram based on univariate and multivariate logistic regression was constructed. The best-performing model and nomogram underwent external validation using the eICU-CRD, and an online risk calculator and a dynamic nomogram visualization tool were developed accordingly.

Results: There were 2,343 HF patients included in the study, while 535 died during hospitalization. The median age was 70 years, with 60.31% being male. The median length of hospital stay was 6 days. Compared to the non-survival group, patients in the survival group were younger (median 69 vs. 73 years),) and had shorter hospital stays. The study included 75 clinical features, encompassing demographic information, vital signs, laboratory tests, hemodynamic parameters, and oxygenation status. Among the 15 machine learning models, LDA demonstrated the best overall performance, with an accuracy of 0.8354 in the training set and 0.8563 in the test set. Additionally, the LDA model achieved favorable results across multiple evaluation metrics, including AUC, recall, precision, F1 score, Cohen's kappa score, and MCC. The concordance index (C-index) was 0.7972 in the training set and 0.8125 in the test set. External validation confirmed that the LDA model maintained strong predictive performance, with accuracy, recall, precision, and F1 score values around 0.9 and an AUC of 0.79. Univariate logistic regression analysis was conducted in the training set, and variables with significant associations were included in the multivariate logistic regression model to develop a nomogram. The final nomogram incorporated 14 key features and demonstrated AUC values of 0.852, 0.855, and 0.770 in the training, test, and external validation sets, respectively. The calibration curve indicated good agreement between predicted and observed outcomes. The study deployed web calculators based on the LDA model and nomogram.

Conclusion: The machine learning model exhibited good predictive performance for 24-hour to 28-day all-cause mortality in hospitalized HF patients, empowering the accuracy and practicality of prognosis assessment. SHAP enhanced the interpretability of machine learning models. The deployed LDA model and dynamic nomogram-based online calculators provide accessible and efficient risk evaluation tools.

Part 3: Stratified analysis of metabolic indicators and blood pressureon cardiovascular risk in middle-aged and older adults

Objective: Cardiovascular disease (CVD) is a major health challenge worldwide, particularly posing a significant threat to older adults. Among potential risk factors, metabolic markers such as the triglyceride-glucose (TyG) index, a key indicator of insulin resistance, may play a role in the evaluation of CVD risk. This study aims to investigate the relationship between metabolic indicators, including TyG, TyG-body mass index (BMI), and their cumulative indices and CVD risk, and further evaluate the impact of blood pressure under different levels of metabolic exposure, to provide guidance for CVD risk evaluation and management.

Methods: This study utilized data from the China Health and Retirement Longitudinal Study (CHARLS), analyzing participants free of CVD at baseline. CHARLS conducts follow-up surveys every two years, with five waves available to date. Baseline data were collected in Wave 1, while incident CVD cases were identified based on follow-up data from Wave 3 and Wave 5. The Cox proportional hazards regression model was employed to examine the associations between TyG, TyG-BMI, cumulative TyG, cumulative TyG-BMI, and CVD risk, adjusting for multiple covariates including age, sex, household registration, education level, marital status, smoking history, hypertension, and kidney disease. Restricted cubic spline regression (RCS) was adopted to explore potential nonlinear relationships between metabolic indicators and CVD risk. A progressive stratified analysis was conducted by classifying participants into low and high metabolic exposure groups based on cumulative TyG and cumulative TyG-BMI, to assess the impact of blood pressure on CVD risk at different metabolic exposure levels.

Results: This retrospective cohort study included 4,800 participants. 55.23% were female and 44.77% were male, with a median age of 57 years old. BMI, TyG, and TyG-BMI significantly increased by Wave 3 compared to baseline. Significant differences in blood pressure, metabolic indicators, and comorbidities were observed between participants who developed CVD and those who did not. Cox proportional hazards regression analysis revealed that TyG, cumulative TyG, TyG-BMI, and cumulative TyG-BMI were significantly associated with long-term CVD risk, with impact increasing over time. The association between TyG and CVD risk was not significant at the 4-year follow-up (HR=1.019, 95% CI 0.891-1.165, P=0.785) but became statistically significant at the 9-year follow-up, with a 13.4% increase in risk (HR=1.134, 95%CI 1.026-1.253, P=0.014). Similarly, cumulative TyG showed a non-significant association at 4 years (HR=1.014, 95%CI 0.962-1.068, P=0.607), but a significant 5.9% risk increase at 9 years (HR=1.059, 95%CI 1.019-1.102, P=0.004). Each 10-unit increase in TyG-BMI was associated with a 3-4% increase in CVD risk, while cumulative TyG-BMI demonstrated a mild effect, with each 10-unit increase leading to a 1-2% risk elevation. When grouped by quartiles, participants in higher quartile of TyG, cumulative TyG, TyG-BMI, and cumulative TyG-BMI exhibited a significantly higher CVD risk than those in the lower quartiles (Q1-Q3). Further RCS regression analysis indicated that TyG and cumulative TyG had a linear positive correlation with CVD risk, the higher the TyG level, the greater the risk of CVD. While TyG-BMI and cumulative TyG-BMI followed an inverted U-shaped pattern, with the greatest impact on CVD risk observed at moderate levels. Stratified analysis based on metabolic exposure showed that elevated blood pressure had a stronger impact on CVD risk under the low metabolic burden. In contrast, under high metabolic burden, the association between blood pressure and CVD risk was weaker and took longer to become apparent. Diastolic blood pressure had a greater impact on CVD risk than systolic blood pressure, regardless of metabolic exposure levels.

Conclusion: TyG and cumulative TyG exhibit a linear positive association with long-term CVD risk in middle-aged and older adults, whereas TyG-BMI and cumulative TyG-BMI follow an inverted U-shaped pattern, with the most evident risk impact at moderate levels. The evaluation value of metabolic indicators for CVD risk strengthens over time. Blood pressure shows a stronger impact on CVD risk under a low metabolic burden, whereas its effect weakens under a high metabolic burden. Diastolic blood pressure has a greater influence on CVD risk than systolic blood pressure.