目的 开发并验证基于床旁肺部超声（lung ultrasound， LUS）与膈肌超声（diaphragm ultrasound， DUS）的可解释机器学习模型，用于术前预测胃肠道肿瘤手术患者的肺功能障碍。 方法 采用前瞻性研究设计，纳入424例患者（2021年6月至2023年12月）用于模型开发，并在一个独立的101例患者队列（2024年1—12月）上进行外部验证。收集了临床变量、PFTs结果和超声参数（LUS评分、膈肌移动度和增厚分数）。使用3种特征选择方法：Least Absolute Shrinkage and Selection Operator （LASSO）、Support Vector Machine Recursive Feature Elimination （SVM-RFE）、eXtreme Gradient Boosting Recursive Feature Elimination （XGBoost-RFE）筛选关键预测因子。使用5折交叉验证训练和评估了5种机器学习算法。基于AUC、准确度、敏感度、特异度、校准和决策曲线分析评估最佳模型。通过SHapley Additive exPlanations（SHAP）分析模型的可解释性。 结果 术前肺功能障碍的患病率为36.8%。确定深呼吸时膈肌移动度［diaphragmatic excursion during deep breathing， D-DE （4.26 cm vs. 5.05 cm， P < 0.001）］、肺部超声评分［LUS score， LUSs （4 vs. 1， P < 0.001）］和深呼吸时膈肌增厚分数［diaphragmatic thickening fraction during deep breathing， D-DTF （39.83% vs. 71.80%， P < 0.001）］为关键预测因子。支持向量机（support vector machine， SVM）模型表现出优异的泛化能力，在内部测试集上AUC达到0.867（准确度80.0%，敏感度81.8%，特异度79.4%）。外部验证证实了其稳健性能，AUC为0.934（95%CI：0.881 ~ 0.987），准确度为88.0%。SHAP分析显示D-DTF是最有影响力的保护性因素。 结论 整合了D-DE、LUSs和D-DTF的可解释SVM模型能够准确预测胃肠道肿瘤患者的术前肺功能障碍，为术前风险评估提供了无创、床旁适用的新方法。

Objective To develop and validate an interpretable machine learning model based on bedside lung and diaphragm ultrasound for preoperative prediction of pulmonary dysfunction in patients undergoing gastrointestinal tumor surgery. Methods In this prospective study， data from 424 patients （June 2021?December 2023） were used for model development， with external validation conducted on an independent cohort of 101 patients （January 2024-December 2024）. Clinical variables， PFTs results， and ultrasound parameters （LUS score， diaphragmatic excursion， and thickening fraction） were collected. Three feature selection methods-Least Absolute Shrinkage and Selection Operator （LASSO）， Support Vector Machine Recursive Feature Elimination （SVM-RFE）， eXtreme Gradient Boosting Recursive Feature Elimination （XGBoost-RFE）- were employed to identify key predictors. Five machine learning algorithms were trained and evaluated using 5-fold cross-validation. The optimal model was assessed based on the area under the receiver operating characteristic curve （AUC）， accuracy， sensitivity， specificity， calibration， and decision curve analysis. SHapley Additive exPlanations （SHAP） analysis was applied to enhance model interpretability. Results The prevalence of preoperative pulmonary dysfunction was 36.8%. Three key predictors were consistently identified： diaphragmatic excursion during deep breathing ［D-DE （4.26 cm vs. 5.05 cm， P < 0.001）］， LUS score ［LUSs （4 vs. 1， P < 0.001）］， and diaphragmatic thickening fraction during deep breathing ［D-DTF （39.83% vs. 71.80%， P < 0.001）］. The Support Vector Machine （SVM） model demonstrated superior generalizability， achieving an AUC of 0.867 on the internal test set （accuracy of 80.0%， sensitivity of 81.8%， specificity of 79.4%）. External validation confirmed its robust performance， with an AUC of 0.934 （95% CI： 0.881 ~ 0.987） and an accuracy of 88.0%. SHAP analysis revealed D-DTF as the most influential protective factor. Conclusion An interpretable SVM model integrating D-DE， LUSs， and D-DTF accurately predicts preoperative pulmonary dysfunction in patients with gastrointestinal tumors， offering a non-invasive， bedside-compatible new approach for preoperative risk assessment.