基于机器学习的胃间质瘤内镜手术术后出血风险预测模型的构建与验证-《中国医学物理学杂志》

文章信息/Info

Title:: Construction and validation of machine learning-based prediction models for postoperativebleeding following endoscopic resection of gastric gastrointestinal stromal tumor

文章编号:: 1005-202X（2025）04-0550-11

作者:: 刘罗杰 1; 陈健 1; 高福利 1; 冯云赋 2; 徐晓丹 1; 1. 苏州大学附属常熟医院（常熟市第一人民医院）消化内科，江苏常熟 215500；2. 昆山市第一人民医院内镜中心，江苏昆山215300

Author(s):: LIU Luojie1; CHEN Jian1; GAO Fuli1; FENG Yunfu2; XU Xiaodan1; 1. Department of Gastroenterology, Changshu No.1 People’s Hospital/Changshu Hospital Affiliated to Soochow University, Changshu215500, China; 2. Endoscopy Center, the First People’s Hospital of Kunshan, Kunshan 215300, China

关键词:: 胃间质瘤; 自动化机器学习; 内镜手术; 术后出血; 预测模型

Keywords:: gastric gastrointestinal stromal tumor; automated machine learning; endoscopic surgery; postoperative bleeding; prediction model

分类号:: R318；R573

DOI:: 10.3969/j.issn.1005-202X.2025.04.018

文献标志码:: A

摘要:: 目的：探讨影响胃间质瘤（gGIST）内镜手术术后出血的危险因素，并应用4种不同机器学习算法构建预测模型，以期准确预测gGIST内镜手术术后出血风险。方法：收集gGIST患者资料，以8：2的比例将研究对象随机分配到训练队列（n=502）和验证队列（n=130）。在训练队列中，应用合成少数类过采样技术（SMOTE）中的变体SMOTE-NC进行过采样。利用梯度提升机（GBM）、深度学习、广义线性模型和分布式随机森林4种机器学习算法构建预测模型。采用最小绝对值收缩和选择算子筛选变量，构建传统逻辑回归（LR）模型。通过计算受试者工作特征曲线下面积（AUC）、灵敏度、特异度、准确度、阳性预测值和阴性预测值评估模型性能。对最优模型进行包括特征重要性、沙普利近似法（SHAP）和力图在内的可解释性分析，并开发一款可实际应用的网络应用程序。结果：在632例患者中，78例（12.3%）发生了术后出血。在验证队列中，对比5种预测模型，GBM模型表现最佳，其AUC值为0.889，95%CI为0.829~0.948，优于其他模型。变量重要性分析显示，术者经验、手术时间、术中大出血、肿瘤大小等因素对预测术后出血具有重要影响。SHAP图和力图展示了变量在二分类预测结果中的分布特征，以及各变量对预测结果的影响。结论：GBM模型对预测gGIST内镜术后出血具有较好的预测价值。同时，网络应用程序的构建方便了临床使用。

Abstract:: Objective To explore the risk factors for postoperative bleeding after endoscopic resection of gastricgastrointestinal stromal tumor (gGIST) and to construct prediction models using 4 different machine learning algorithms foraccurately predicting postoperative bleeding. Methods The clinical data of gGIST patients were collected, and the patientswere randomly divided into a training cohort (n=502) and a validation cohort (n=130) at an 8:2 ratio. Synthetic minority oversampling technique-nominal continuous was used for oversampling in the training cohort. Four prediction models wereconstructed using gradient boost machine (GBM), deep learning, generalized linear model and distributed random forest,separately; and in addition, the least absolute shrinkage and selection operator was used to screen variables and construct atraditional Logistic regression model. Model performance was evaluated by calculating the area under the receiver operatingcharacteristic curve (AUC), sensitivity, specificity, accuracy, positive predictive value and negative predictive value.Interpretability analyses, including feature importance, SHapley additive exPlanation and force plot, were performed on theoptimal model, and a practically applicable web application was developed. Results Among 632 patients, 78 (12.3%)experienced postoperative bleeding. In the validation cohort, GBM model performed best among 5 prediction models, with an AUC value of 0.889 and a 95%CI of 0.829-0.948, superior to the other 4 models. Variable importance analysis identifiedsurgeon experience, operation time, intraoperative hemorrhage, tumor size as the factors affecting postoperative bleedingprediction. The SHapley additive exPlanation plot and force plot showed the distribution characteristics of variables in thebinary classification prediction results and the effect of each variable on the prediction results. Conclusion GBM model hashigh predictive value for postoperative bleeding following endoscopic resection of gGIST, and the construction of the webapplication facilitates its clinical use.

《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]

文章信息/Info

备注/Memo

常用功能

导航/Navigate

工具/Tools

统计/Statistics