Page 17 - 《南京医科大学学报》自然科学版2026年第2期

第46卷第2期 南京医科大学学报（自然科学版）
2026年2月 Journal of Nanjing Medical University（Natural Sciences） ·173 ·

·专题研究：心脏疾病·

基于机器学习的心脏术后衰弱预测模型的构建与验证

杨 玥，葛 愿，李明辉，耿丹丹 *
南京医科大学第一附属医院心脏大血管外科，江苏 南京 210029

［摘 要］ 目的：构建并验证一种用于预测心脏术后衰弱风险的个体化预测模型。方法：纳入2023年1—12月在南京医科大
学第一附属医院接受心脏手术的患者，在术后1个月采用衰弱筛查量表对患者进行衰弱评估。根据评估结果将患者分为衰弱
组及非衰弱组。采用最小绝对收缩和选择算子（least absolute shrinkage and selection operator，LASSO）、随机森林（random forest，
RF）以及极限梯度提升（extreme gradient boosting，XGBoost）3 种机器学习算法筛选共同预测因子，随后使用 Logistic回归构建
列线图模型。采用受试者工作特征（receiver operating characteristic，ROC）曲线及其曲线下面积（area under the curve，AUC）评估
模型的区分度、校准曲线评估模型的一致性、决策曲线分析（decision curve analysis，DCA）评估该模型的临床价值，并在内部验证
集以及时间分层验证集中予以验证。结果：共纳入301例患者，其中235例按7∶3比例分为训练集（n=165）、内部验证集（n=70），
其余 66 例患者作为时间分层验证。依据机器学习结果，纳入 4 个共同预测因子：年龄、术前左心室射血分数（left ventricular
ejection fraction，LVEF）、术前白蛋白水平和术前左心室舒张末期内径（left ventricular diastolic dimension，LVDd）。以此构建列线
图，在训练集（AUC=0.846，95%CI：0.763~0.928）、内部验证集（AUC=0.821，95%CI：0.701~0.940）和时间分层验证集（AUC=0.846，
95%CI：0.740~0.951）中均表现出优异的区分能力。校准曲线显示预测风险与观察风险之间具有高度一致性。DCA进一步证
明了其良好的临床实用性。结论：基于患者年龄、术前白蛋白水平、LVEF以及LVDd构建的心脏术后衰弱预测列线图模型具
有良好的预测效能与临床适用性，有助于早期识别高危患者。
［关键词］ 机器学习；预测模型；心脏手术；衰弱；列线图
［中图分类号］ R619 ［文献标志码］ A ［文章编号］ 1007⁃4368（2026）02⁃173⁃09
doi：10.7655/NYDXBNSN251184

Development and validation of a machine learning⁃based nomogram model for predicting
frailty after cardiac surgery
*
YANG Yue，GE Yuan，LI Minghui，GENG Dandan
Department of Cardiovascular Surgery，the First Affiliated Hospital of Nanjing Medical University，Nanjing 210029，
China

［Abstract］ Objective：To develop and externally validate a patient ⁃ level prediction model for postoperative frailty in adults
undergoing cardiac surgery. Methods：Our study included patients who underwent cardiac surgery at the First Affiliated Hospital of
Nanjing Medical University between January and December 2023. Frailty was assessed one month postoperatively using the FRAIL
scale，and patients were categorized into frailty and non⁃frailty groups. Three machine learning algorithms，least absolute shrinkage and
selection operator（LASSO），random forest（RF），and extreme gradient boosting（XGBoost），were employed to identify common
predictors. A nomogram prediction model was subsequently constructed using Logistic regression. The model’s discriminative ability
was evaluated using the receiver operating characteristic（ROC）curve and the area under the curve（AUC）. Calibration curves
assessed consistency，and decision curve analysis（DCA）evaluated clinical utility. The model was validated internally and externally.
Results：A total of 301 patients were included. Among them，235 patients were divided into a training set（n=165）and an internal
validation set（n=70）at a ratio of 7∶3，while the remaining 66 patients served as temporal validation. Based on the machine learning
results，four common predictors were identified：age，preoperative left ventricular ejection fraction（LVEF），preoperative albumin level，
and preoperative left ventricular end ⁃ diastolic dimension（LVDd）. These were used to construct the nomogram. The model

［基金项目］ 国家自然科学基金（82200340）；中国社会福利基金会护爱基金（HLCXKT⁃20230180）
通信作者（Corresponding author），E⁃mail：15850603835@163.com（ORCID：0009⁃0002⁃2863⁃8623）
∗