基于机器学习构建COPD合并心血管疾病患者发生死亡的风险预测模型

A Machine Learning-Based Risk Prediction Model for Mortality in Patients with COPD Complicated by Cardiovascular Disease

摘要: 目的：为了分析4种机器学习算法构建慢性阻塞性肺疾病(COPD)合并心血管疾病发生全因死亡率预测模型的预测价值。方法：从NHANES数据库中提取2013~2018年三个调查周期的数据，依据COPD合并心血管疾病的诊断进行筛选，纳入490名研究对象。跨周期整合数据，对数据缺失、极端值及不合理记录进行处理，对变量进行筛选后构建4种机器学习算法，以全因死亡率为结局指标，对慢性阻塞性肺疾病合并心血管疾病患者发生全因死亡的风险进行预测模型构建，并且比较4种预测模型，并通过ROC曲线、校准曲线及DCA方法对模型性能进行系统评价。结果：经单因素、LASSO回归筛选变量分析，共筛选出10项具有预测意义的变量，包括吸烟、血红蛋白、谷草转氨酶、淋巴细胞计数、体重指数、单核细胞计数、年龄、免疫炎症指数、低密度脂蛋白及总胆固醇。根据筛选的变量，使用SVM、Logistic回归、RF及XGBoost进行预测模型构建，在训练集中，随机森林(Random Forest, RF)、支持向量机(SVM)、极限梯度提升(XGBoost)和Logistic回归模型的ROC曲线下面积分别为0.8353、0.8571、0.8289和0.7573。在测试集中，RF、SVM、XGBoost和Logistic回归模型的AUC分别为0.7277、0.7294、0.6897和0.7220。结论：四种模型在训练集区分能力均较好，AUC最高的模型是SVM模型，预测性能最佳，在测试集中LR与SVM的AUC接近，提示LR具有相当的泛化判别能力。测试集与训练集相比，各模型AUC均有所下降，但整体仍保持中等以上的预测效能。综合训练集与测试集ROC曲线分析结果，SVM算法和LR算法构建的COPD合并心血管疾病患者死亡风险预测模型具有最优的判别能力和较好的泛化性能，可作为一种有效工具，用于早期识别COPD合并心血管疾病患者发生死亡风险，为临床风险分层和干预决策提供参考依据。

Abstract: Objective: To analyze the predictive value of four machine learning algorithms in constructing all-cause mortality prediction models for patients with chronic obstructive pulmonary disease (COPD) and cardiovascular disease. Methods: Data from three survey periods (2013~2018) were extracted from the NHANES database. Participants were screened based on their diagnosis of COPD combined with cardiovascular disease, and 490 subjects were included. Data across periods were integrated, and missing data, outliers, and inappropriate records were removed. After variable screening, four machine learning algorithms were constructed to predict the risk of all-cause mor-tality in patients with COPD and cardiovascular disease, using all-cause mortality as the outcome indicator. The four prediction models were compared, and their performance was systematically evaluated using ROC curves, calibration curves, and DCA methods. Results: Univariate and LASSO regression analyses identified 10 predictive variables: smoking status, hemoglobin, aspartate ami-notransferase (AST), lymphocyte count, body mass index (BMI), monocyte count, age, im-mune-inflammatory index, low-density lipoprotein (LDL), and total cholesterol. Using these selected variables, prediction models were constructed with RF, SVM, logistic regression, and XGBoost. On the training set, the areas under the ROC curves for Random Forest (RF), Support Vector Machine (SVM), XGBoost, and logistic regression were 0.8353, 0.8571, 0.8289, and 0.7573, respectively. On the test set, the AUC values for RF, SVM, XGBoost, and Logistic regression were 0.7277, 0.7294, 0.6897, and 0.7220, respectively. Conclusion: All four models demonstrated good discrimination capabilities on the training set, with the SVM model achieving the highest AUC and best predictive performance. In the test set, the AUCs of LR and SVM were similar, suggesting that LR has compara-ble generalization and discriminative ability. Compared to the training set, the AUC of each model decreased on the test set, yet overall maintained above-average predictive efficacy. Based on the combined ROC curve analysis of the training and test sets, the SVM- and LR-based model for pre-dicting mortality risk in COPD patients with CVD demonstrates optimal discriminative capability and good generalization performance. It serves as an effective tool for early identification of mortal-ity risk in COPD patients with cardiovascular disease, providing a reference for clinical risk stratifi-cation and intervention decisions.