Hurley Nathan C, Desai Nihar, Dhruva Sanket S, Khera Rohan, Schulz Wade, Huang Chenxi, Curtis Jeptha, Masoudi Frederick, Rumsfeld John, Negahban Sahand, Krumholz Harlan M, Mortazavi Bobak J
Department of Computer Science & Engineering, Texas A&M University, College Station, Texas, United States of America.
Center for Outcomes Research and Evaluation, Yale New Haven Health, New Haven, Connecticut, United States of America.
PLOS Digit Health. 2025 Jun 25;4(6):e0000906. doi: 10.1371/journal.pdig.0000906. eCollection 2025 Jun.
While static risk models may identify key driving risk factors, the dynamic nature of risk requires up-to-date risk information to guide treatment decision making. Bleeding is a complication of percutaneous coronary intervention (PCI), and existing risk models produce only a single risk estimate anchored at a single point in time, despite the dynamic nature of this risk. Using data available from the National Cardiovascular Data Registry (NCDR) CathPCI, we trained 6 different tree-based machine learning models to estimate the risk of bleeding at key decision points: 1) choice of access site, 2) prescription of medication before PCI, and 3) choice of closure device. We began with 3,423,170 PCIs performed between July 2009 through April 2015. We included only index PCIs and removed anyone who had missing data regarding bleeding events or underwent coronary artery bypass grafting during the index admission. We included 2,868,808 PCIs; 2,314,446 (80.7%) before 2014 for training and 554,362 (19.3%) remaining for validation. This study considered all data available from the Registry prior to patient discharge: patient characteristics, coronary anatomy and lesion characterization, laboratory data, past medical history, anti-coagulation, stent type, and closure method categories. The primary outcome was any in-hospital bleeding event within 72 hours after the start of the PCI procedure. Discrimination improved from an area under the receiver operating characteristic curve (AUROC) of 0.812 using only presentation variables to 0.845 using all variables. Among 123,712 patients classified as low risk by the initial model, 14,441 were reclassified as moderate risk (1.4% experienced bleeds), while 723 were reclassified as high risk (12.5% experienced bleeds). Static risk prediction models have more predictive error than those that update risk prediction with newly available data, which provides up-to-date risk prediction for individualized care throughout a hospitalization.
虽然静态风险模型可以识别关键的驱动风险因素，但风险的动态性质需要最新的风险信息来指导治疗决策。出血是经皮冠状动脉介入治疗（PCI）的一种并发症，尽管这种风险具有动态性质，但现有的风险模型仅产生一个基于单一时间点的单一风险估计值。利用从国家心血管数据注册库（NCDR）CathPCI获得的数据，我们训练了6种不同的基于树的机器学习模型，以估计在关键决策点的出血风险：1）入路部位的选择，2）PCI术前药物处方，以及3）闭合装置的选择。我们从2009年7月至2015年4月期间进行的3423170例PCI手术开始。我们仅纳入首次PCI手术，并排除了任何在首次住院期间有出血事件缺失数据或接受冠状动脉旁路移植术的患者。我们纳入了2868808例PCI手术；2014年之前的2314446例（80.7%）用于训练，其余554362例（19.3%）用于验证。本研究考虑了患者出院前注册库中的所有可用数据：患者特征、冠状动脉解剖结构和病变特征、实验室数据、既往病史、抗凝情况、支架类型和闭合方法类别。主要结局是PCI手术开始后72小时内的任何院内出血事件。辨别能力从仅使用就诊变量时的受试者工作特征曲线下面积（AUROC）0.812提高到使用所有变量时的0.845。在初始模型分类为低风险的123712例患者中，14441例被重新分类为中度风险（1.4%发生出血），而723例被重新分类为高风险（12.5%发生出血）。静态风险预测模型比那些使用新获得的数据更新风险预测的模型具有更多的预测误差，后者可为整个住院期间的个体化护理提供最新的风险预测。
