摘要目的:探讨极限梯度提升(XGBoost)算法分析急诊经皮冠状动脉介入治疗(PCI)术后心电图特征预测院内主要不良心血管事件(MACE)的价值。方法:采取回顾性研究，抽取2024年9月至2025年9月于安阳市人民医院接受急诊PCI治疗的236例AMI患者，按是否发生院内MACE分为MACE组(43例)和非MACE组(193例)。比较两组临床资料及术后心电图特征指标；应用R(R4.1.0)软件包通过XGBoost算法筛选影响因素，用二元Logistic回归分析检验患者发生院内MACE的影响因素；采用受试者工作特征曲线分析XGBoost模型预测患者发生院内MACE的价值；采用决策曲线评估基于XGBoost算法构建的院内MACE预测模型的临床净获益。结果:MACE组置入支架数≥2个占比、微伏级T波电交替(MTWA)高于非MACE组( P<0.05)，MACE组左室射血分数(LVEF)、RR间期的标准差(SDNN)低于非MACE组( P<0.05)。基于XGBoost算法筛选出Gain>0.15的高贡献因素共3个，分别是MTWA(Gain＝0.29)、LVEF(Gain＝0.19)、SDNN(Gain＝0.17)，经二元Logistic回归分析结果显示，LVEF、SDNN是患者发生院内MACE的保护因素( OR均<1， P均<0.05)，MTWA是患者发生院内MACE的危险因素( OR>1， P<0.05)。采用5折交叉验证策略评估模型性能，XGBoost模型经网格搜索优化后，确定最优超参数组合为：学习率＝0.05，最大树深度＝5，子采样率＝0.8，正则化参数＝1。基于该模型计算的特征重要性(Gain值)显示，MTWA(Gain＝0.29)、LVEF(Gain＝0.19)和SDNN(Gain＝0.17)为前三位高贡献变量(Gain>0.15)，故将其纳入最终预测模型。经5折交叉验证，XGBoost模型预测院内MACE的曲线下面积为0.791(95% CI: 0.732~0.846)，平均准确率为70.00%，敏感度为0.462，特异度为0.754，F1分数为0.364。校准曲线显示模型预测概率与实际事件发生率基本一致(Hosmer-Lemeshow检验 χ2＝8.06， P＝0.427)，提示模型校准良好，Brier得分为0.18，进一步证实模型预测校准度可接受；决策曲线分析结果显示，联合模型在阈值概率为25%时达到最大净获益。 结论:急诊PCI术后患者的动态心电图指标中MTWA、SDNN，及心脏功能指标LVEF是预测院内MACE的独立危险因素，基于XGBoost算法构建的预测模型具有良好的预测效能，可为临床早期识别高危患者、制定个体化干预策略提供参考依据。

abstractsObjective:To investigate the value of eXtreme Gradient Boosting (XGBoost) algorithm in analyzing electrocardiographic characteristics after emergency percutaneous coronary intervention (PCI) for predicting in-hospital major adverse cardiovascular events (MACE).Methods:A retrospective study was conducted to select 236 patients with acute myocardial infarction (AMI) who underwent emergency PCI at Anyang People’s Hospital from September 2024 to September 2025. Patients were divided into the MACE group (43 cases) and non-MACE group (193 cases) according to the occurrence of in-hospital MACE. The clinical data and postoperative electrocardiographic indicators were compared between the two groups. The R software (version 4.1.0) package was used to screen influencing factors via the XGBoost algorithm, and binary logistic regression analysis was performed to verify factors associated with in-hospital MACE. Receiver operating characteristic curve was applied to evaluate the predictive value of the XGBoost model for in-hospital MACE, and decision curve analysis was used to assess the clinical net benefit of the XGBoost-based prediction model for in-hospital MACE.Results:The proportion of patients with ≥2 stents implanted and microvolt T-wave alternans (MTWA) were higher in the MACE group than in the non-MACE group ( P<0.05), while left ventricular ejection fraction (LVEF) and standard deviation of normal-to-normal RR intervals (SDNN) were lower in the MACE group than in the non-MACE group ( P<0.05). Three high-contribution factors with Gain > 0.15 were screened by the XGBoost algorithm: MTWA (Gain=0.29), LVEF (Gain=0.19), and SDNN (Gain=0.17). Binary logistic regression analysis showed that LVEF and SDNN were protective factors for in-hospital MACE (both OR<1, both P<0.05), whereas MTWA was a risk factor ( OR>1, P<0.05). The model performance was evaluated using 5-fold cross-validation. After hyperparameter optimization through grid search, the optimal hyperparameter combination for the XGBoost model was determined as follows: learning rate=0.05, maximum tree depth=5, subsample ratio=0.8, and regularization parameter=1. The feature importance (Gain values) calculated based on this model identified that MTWA (Gain=0.29), LVEF (Gain=0.19), and SDNN (Gain=0.17) were the top three high-contribution variables (Gain>0.15), and thus were included in the final prediction model. With 5-fold cross-validation, the area under the curve of the XGBoost model for predicting in-hospital MACE was 0.791 (95% CI: 0.732-0.846), with an average accuracy of 70.00%, sensitivity of 0.462, specificity of 0.754, and F1-score of 0.364. The calibration curve showed that the predicted probabilities were generally consistent with the observed event rates (Hosmer-Lemeshow test χ2=8.06, P=0.427), indicating satisfactory model calibration; and the Brier score was 0.18, further confirming acceptable predictive calibration. Decision curve analysis indicated that the combined model achieved the maximum net benefit at a threshold probability of 25%. Conclusions:Among ambulatory electrocardiographic parameters and cardiac function indicators in patients after emergency PCI, MTWA, SDNN, and LVEF are independent predictors of in-hospital MACE. The prediction model constructed based on the XGBoost algorithm shows favorable predictive performance, which can provide a reference for early identification of high-risk patients and formulation of individualized intervention strategies in clinical practice.