Xu Man, Liu Hao, Dai Anran, Tan Qilian, Zhang Xinlong, Ding Rui, Chen Chen, Zou Jianjun, Li Yongjun, Si Yanna
Department of Anesthesiology, Perioperative and Pain Medicine, Nanjing First Hospital, Nanjing Medical University, Nanjing, China.
State Key Laboratory of Natural Medicines, Key Laboratory of Drug Metabolism, China Pharmaceutical University, Nanjing, China.
BMC Anesthesiol. 2025 Aug 5;25(1):394. doi: 10.1186/s12871-025-03239-z.
Postoperative respiratory failure following cardiac surgery (CS-PRF) remains a critical complication with substantial morbidity and mortality. Current risk prediction models are limited by static assessments and suboptimal accuracy. This study aimed to develop and validate a dynamic, machine learning–based model to enhance perioperative risk stratification for CS-PRF.
We retrospectively analyzed 1,016 adult patients who underwent cardiac surgery. Feature selection was conducted via the Least Absolute Shrinkage and Selection Operator (LASSO) and Boruta algorithms. Five machine learning models, including logistic regression, multilayer perceptron, extreme gradient boosting, categorical boosting, and deep neural network (DNN), were trained using preoperative and intraoperative variables. Model performance was evaluated by the area under the receiver operating characteristic curve (AUROC), area under the precision–recall curve (AUPRC), and calibration metrics. Model interpretability was evaluated via SHapley additive exPlanation (SHAP), and restricted cubic spline (RCS) analyses were used to explore nonlinear associations.
The incidence of CS-PRF was 16.3%. In the validation cohort, the DNN model achieved superior performance, with an AUROC of 0.782 (95% CI 0.703–0.852) and an AUPRC of 0.496 based on preoperative variables, which improved to an AUROC of 0.855 (95% CI 0.796–0.906) and an AUPRC of 0.549 with the addition of intraoperative data. Calibration analysis demonstrated good agreement between predicted and observed risk. SHAP analysis of the preoperative model identified pulmonary artery pressure, age, and preoperative creatinine as key contributors. In the combined model, intraoperative features such as cardiopulmonary bypass duration and autologous blood transfusion volume emerged as additional important predictors. RCS analysis revealed a nonlinear association between age and CS-PRF. A web-based risk calculator integrating DNN predictions and individualized SHAP interpretation was deployed to support clinical decision-making.
We developed a deep learning model that integrates perioperative variables to predict postoperative respiratory failure following cardiac surgery. Demonstrating high accuracy and interpretability, the model has been deployed as an accessible web-based calculator, offering a practical tool for personalized perioperative risk assessment.
Not applicable. This study is a retrospective observational study and was not registered as a clinical trial.
The online version contains supplementary material available at 10.1186/s12871-025-03239-z.
心脏手术后的呼吸衰竭（CS-PRF）仍然是一种严重的并发症，具有较高的发病率和死亡率。目前的风险预测模型受限于静态评估和欠佳的准确性。本研究旨在开发并验证一种基于机器学习的动态模型，以加强对CS-PRF的围手术期风险分层。
我们回顾性分析了1016例接受心脏手术的成年患者。通过最小绝对收缩和选择算子（LASSO）算法和博鲁塔算法进行特征选择。使用术前和术中变量对包括逻辑回归、多层感知器、极端梯度提升、分类提升和深度神经网络（DNN）在内的五种机器学习模型进行训练。通过受试者操作特征曲线下面积（AUROC）、精确召回率曲线下面积（AUPRC）和校准指标评估模型性能。通过SHapley加性解释（SHAP）评估模型可解释性，并使用受限立方样条（RCS）分析探索非线性关联。
CS-PRF的发生率为16.3%。在验证队列中，DNN模型表现优异，基于术前变量的AUROC为0.782（95%CI 0.703–0.852），AUPRC为0.496，加入术中数据后，AUROC提高到0.855（95%CI 0.796–0.906），AUPRC提高到0.549。校准分析表明预测风险与观察到的风险之间具有良好的一致性。术前模型的SHAP分析确定肺动脉压、年龄和术前肌酐为关键因素。在联合模型中，诸如体外循环持续时间和自体输血量等术中特征成为另外的重要预测因素。RCS分析揭示年龄与CS-PRF之间存在非线性关联。部署了一个整合DNN预测和个性化SHAP解释的基于网络的风险计算器，以支持临床决策。
我们开发了一种整合围手术期变量的深度学习模型，用于预测心脏手术后的呼吸衰竭。该模型具有高准确性和可解释性，已作为一个易于访问的基于网络的计算器进行部署，为个性化围手术期风险评估提供了一个实用工具。
不适用。本研究是一项回顾性观察性研究，未注册为临床试验。
在线版本包含可在10.11
