多算法机器学习模型在胰十二指肠切除术后胰瘘风险预测中的构建与比较

Construction and comparison of multi-algorithm machine learning models for predicting postoperative pancreatic fistula after pancreaticoduodenectomy

摘要:

背景与目的 胰十二指肠切除术（PD）后临床相关性术后胰瘘（CR-POPF）是影响患者预后的关键并发症之一。传统风险评估模型在复杂非线性关系处理方面存在局限。本研究旨在筛选CR-POPF的独立危险因素，并基于多种机器学习算法构建与比较预测模型，以提高风险评估的准确率。方法 回顾性纳入2016年1月—2025年12月东南大学附属中大医院行PD的334例患者。采用单因素及多因素Logistic回归筛选CR-POPF独立危险因素。按照7∶3比例随机分为训练集和验证集，基于Logistic回归（LR）、人工神经网络（ANN）、决策树（DT）、随机森林（RF）及支持向量机（SVM）构建预测模型。通过受试者工作特征曲线下面积（AUC）、敏感度、特异度、阳性预测值、阴性预测值、F1值及准确度评估模型性能。结果 多因素分析显示，体质量指数升高（OR=1.167）、糖尿病（OR=3.826）、高血压（OR=2.232）、腹部手术史（OR=2.599）、术前白蛋白降低（OR=0.625）、术后白细胞升高（OR=1.091）及胰腺来源病变（OR=2.945）为CR-POPF的独立危险因素（均P<0.05）。模型比较结果显示，ANN模型表现最佳，其AUC为0.866，敏感度0.745，特异度0.914，F1值0.768，整体预测效能优于其他模型。结论 PD术后CR-POPF的发生与多种临床因素相关。基于ANN构建的预测模型具有较高判别能力，可用于高危患者的早期识别及个体化干预，对提高手术安全性和改善预后具有潜在临床价值。

Abstract:

Background and Aims Clinically relevant postoperative pancreatic fistula (CR-POPF) remains a major complication after pancreaticoduodenectomy (PD), significantly affecting patient outcomes. Conventional risk models have limitations in capturing complex nonlinear relationships. This study aimed to identify independent risk factors for CR-POPF and to develop and compare multiple machine learning-based prediction models.Methods A total of 334 patients who underwent PD at the Hepatobiliary and Pancreatic Center of Zhongda Hospital, Southeast University between January 2016 and December 2025 were retrospectively analyzed. Independent risk factors were identified using univariate and multivariate logistic regression analyses. The dataset was randomly divided into training and validation sets at a 7∶3 ratio. Prediction models were developed using Logistic regression (LR), artificial neural network (ANN), decision tree (DT), random forest (RF), and support vector machine (SVM). Model performance was evaluated using AUC, sensitivity, specificity, positive predictive value, negative predictive value, F1-score, and accuracy.Results Multivariate analysis identified increased BMI (OR=1.167), diabetes mellitus (OR=3.826), hypertension (OR=2.232), history of abdominal surgery (OR=2.599), lower preoperative albumin (OR=0.625), elevated postoperative white blood cell count (OR=1.091), and pancreatic-origin lesions (OR=2.945) as independent risk factors for CR-POPF (all P<0.05). Among the models, the ANN model demonstrated superior performance, with an AUC of 0.866, sensitivity of 0.745, specificity of 0.914, and F1-score of 0.768.Conclusion CR-POPF after PD is influenced by multiple clinical factors. The ANN-based model shows strong predictive performance and may serve as a valuable tool for early identification of high-risk patients and implementation of individualized interventions.