Objective: To construct an ensemble machine learning model for predicting the occurrence of clinically relevant postoperative pancreatic fistula (CR-POPF) after pancreaticoduodenectomy and evaluate its application value. Methods: This is a research on predictive model. Clinical data of 421 patients undergoing pancreaticoduodenectomy in the Department of Pancreatic Surgery,Union Hospital, Tongji Medical College,Huazhong University of Science and Technology from June 2020 to May 2023 were retrospectively collected. There were 241 males (57.2%) and 180 females (42.8%) with an age of (59.7±11.0)years (range: 12 to 85 years).The research objects were divided into training set (315 cases) and test set (106 cases) by stratified random sampling in the ratio of 3∶1. Recursive feature elimination is used to screen features,nine machine learning algorithms are used to model,three groups of models with better fitting ability are selected,and the ensemble model was constructed by Stacking algorithm for model fusion. The model performance was evaluated by various indexes,and the interpretability of the optimal model was analyzed by Shapley Additive Explanations(SHAP) method. The patients in the test set were divided into different risk groups according to the prediction probability (P) of the alternative pancreatic fistula risk score system (a-FRS). The a-FRS score was validated and the predictive efficacy of the model was compared. Results: Among 421 patients,CR-POPF occurred in 84 cases (20.0%). In the test set,the Stacking ensemble model performs best,with the area under the curve (AUC) of the subject's work characteristic curve being 0.823,the accuracy being 0.83,the F1 score being 0.63,and the Brier score being 0.097. SHAP summary map showed that the top 9 factors affecting CR-POPF after pancreaticoduodenectomy were pancreatic duct diameter,CT value ratio,postoperative serum amylase,IL-6,body mass index,operative time,albumin difference before and after surgery,procalcitonin and IL-10. The effects of each feature on the occurrence of CR-POPF after pancreaticoduodenectomy showed a complex nonlinear relationship. The risk of CR-POPF increased when pancreatic duct diameter<3.5 mm,CT value ratio<0.95,postoperative serum amylase concentration>150 U/L,IL-6 level>280 ng/L,operative time>350 minutes,and albumin decreased by more than 10 g/L. The AUC of a-FRS in the test set was 0.668,and the prediction performance of a-FRS was lower than that of the Stacking ensemble machine learning model. Conclusion: The ensemble machine learning model constructed in this study can predict the occurrence of CR-POPF after pancreaticoduodenectomy,and has the potential to be a tool for personalized diagnosis and treatment after pancreaticoduodenectomy.
目的： 构建并验证预测胰十二指肠切除术后患者发生临床相关术后胰瘘（CR-POPF）的集成机器学习模型。 方法： 本研究为预测模型研究。回顾性收集2020年6月至2023年5月在华中科技大学同济医学院附属协和医院胰腺外科接受胰十二指肠切除术的421例患者的临床资料。其中男性241例（57.2%），女性180例（42.8%），年龄（59.7±11.0）岁（范围：12~85岁）。通过分层随机抽样法将研究对象按照3∶1的比例分为训练集（315例）和测试集（106例）。使用递归特征消除算法对特征进行筛选，运用9种机器学习算法分别建模，挑选拟合能力较优的三组模型，通过Stacking算法进行模型融合构建集成模型。通过多种指标评估模型性能，并使用Shapley Additive Explanations（SHAP）方法对最优模型进行可解释性分析。根据替代胰瘘风险评分系统（a-FRS）的预测概率（P）将测试集患者分为不同风险组，对a-FRS评分进行验证并比较其与所建模型的预测效能。 结果： 421例患者中，发生CR-POPF 84例（20.0%）。在测试集中，Stacking集成模型表现最佳，其受试者工作特征曲线的曲线下面积（AUC）为 0.823，准确率为0.83，F1分数为0.63，Brier 得分为0.097。SHAP总结图显示，影响胰十二指肠切除术后发生CR-POPF的前9位因素依次为胰管直径、CT值比值、术后血清淀粉酶、IL-6水平、体重指数、手术时间、术前术后白蛋白差值、降钙素原及IL-10。各个因素对胰十二指肠切除术后CR-POPF 发生的影响均呈现出复杂的非线性关系。当胰管直径<3.5 mm、CT 值比值<0.95、术后血清淀粉酶浓度>150 U/L、IL-6 水平>280 ng/L、手术时间>350 min、白蛋白降低超过10 g/L时，发生CR-POPF的风险增加。a-FRS在测试集中的AUC为0.668，预测效能低于Stacking集成机器学习模型。 结论： 本研究构建的Stacking集成机器学习模型能够预测胰十二指肠切除术后CR-POPF的发生，有潜力成为胰十二指肠切除术后个性化诊疗的有效工具。.