目的·急性心肌梗死（acute myocardial infarction，AMI）患者在接受经皮冠状动脉介入治疗（percutaneous coronary intervention，PCI）时，可能会出现低血压、恶性心律失常等严重并发症，这些并发症会直接影响手术的成功率以及患者的预后。在当前的临床评估中，缺乏针对性的工具，而传统的评估工具又存在预测效能不足的局限性。本研究创新性地运用机器学习方法，旨在构建一个精准的术中并发症预测模型，以解决现有评分系统对非线性关系捕捉不足的问题，为术者提供一个风险预警工具。方法·研究纳入了2019年至2022年上海市奉贤区中心医院就诊的811例急诊PCI患者，并定义了53项候选变量。采用多阶段特征工程框架，包括单变量初筛、Spearman等级相关性验证（Spearman rank correlation analysis）以及SHAP稳定性优化（SHapley additive exPlanations stability optimization），最终通过极端梯度提升（eXtreme gradient boosting，XGBoost）算法构建了预测模型。主要终点为术中复合并发症，即低血压、恶性心律失常以及严重缓慢型心律失常。结果·模型筛选出了6项核心预测因子，分别是急诊左心室射血分数（left ventricular ejection fraction，LVEF）最低值、罪犯血管为右冠状动脉（culprit vessel in right coronary artery，CVRCA）、存在罪犯血管（culprit vessel presence，CVP）、罪犯血管为左前降支动脉（culprit vessel in left anterior descending artery，CVADA）、B型利钠肽（B-type natriuretic peptide，BNP）和心率（heart rate，HR）。XGBoost模型在训练集与验证集的曲线下面积（area under the curve，AUC）分别为0.88和0.84。结论·成功构建了基于XGBoost算法的AMI患者PCI术中并发症预测模型，其预测效能显著优于传统评分系统。该模型通过自动筛选关键临床特征，有效捕捉变量间的复杂交互作用，能够为术者提供个体化的风险评估，从而更好地指导临床决策，提高手术成功率和患者预后。

Objective ·Patients with acute myocardial infarction (AMI) undergoing percutaneous coronary intervention (PCI) are at risk of severe complications, such as hypotension and malignant arrhythmias, which directly affect procedural success and patient prognosis. Current clinical practice lacks targeted assessment tools, and traditional assessment tools have limitations in predictive efficacy. This study innovatively applies machine learning methods to construct a precise intraoperative complication prediction model, addressing the insufficient non-linear relationship capture in existing scoring systems and providing an early risk warning tool for surgeons. Methods ·The study included 811 emergency PCI patients who were treated at Shanghai Fengxian Central Hospital from 2019 to 2022, and defined 53 candidate variables. A multi-stage feature engineering framework was employed, including univariate screening, Spearman rank correlation analysis verification, and SHapley Additive exPlanations (SHAP)-based stability optimization. The prediction model was ultimately constructed using the eXtreme Gradient Boosting (XGBoost) algorithm.The primary endpoint was a composite of intraoperative complications, including hypotension, malignant arrhythmias, and severe bradyarrhythmias. Results ·The model identified six core predictive factors: lowest left ventricular ejection fraction (LVEF), culprit vessel in the right coronary artery (CVRCA), culprit vessel presence (CVP), culprit vessel in the left anterior descending artery (CVADA), B-type natriuretic peptide (BNP), and heart rate (HR). The XGBoost model achieved areas under the curve (AUCs) of 0.88 and 0.84 in the training and validation sets, respectively. Conclusion ·This study successfully constructed, for the first time, an intraoperative complication prediction model for AMI patients undergoing PCI based on the XGBoost algorithm. Its predictive performance significantly outperforms that of traditional scoring systems. By automatically selecting key clinical features and effectively capturing complex interactions between variables, the model provides individualized risk assessment for surgeons, thereby supporting clinical decision-making and potentially improving procedural success rates and patient prognosis.