基于临床和CT影像学特征的胰腺囊性病变进展预测模型的建立及验证  

Establishment and validation of a predictive model for the progression of pancreatic cystic lesions based on clinical and CT radiological features

作　　者：邓文祎 谢飞扬 毛丽 李秀丽 孙照勇[1] 徐凯[1] 朱亮[1] 金征宇[1] 李骁 薛华丹[1] Deng Wenyi;Xie Feiyang;Mao Li;Li Xiuli;Sun Zhaoyong;Xu Kai;Zhu Liang;Jin Zhengyu;Li Xiao;Xue Huadan(Department of Radiology,Peking Union Medical College Hospital,Chinese Academy of Medical Sciences and Peking Union Medical College,Beijing 100730,China;AI Lab,Deepwise Healthcare,Beijing 100080,China;Department of Medical Imaging,Jinling Hospital,Affiliated Hospital of Medical School,Nanjing University,Nanjing 210002,China)

机构地区：[1]中国医学科学院北京协和医学院、北京协和医院放射科,北京100730 [2]深睿医疗人工智能研究院,北京100080 [3]南京大学医学院附属金陵医院(东部战区总医院)放射诊断科,南京210002

出　　处：《中华胰腺病杂志》2024年第1期23-28,共6页Chinese Journal of Pancreatology

基　　金：国家自然科学基金(82071896);国家重点研发计划资助(2020YFC2002702);中央高水平医院临床科研业务费(2022-PUMCH-B-069)。

摘　　要：目的构建基于临床及CT影像学特征的胰腺囊性病变(PCLs)进展预测机器学习模型,并在内、外部测试集上评估其预测效能。方法回顾性收集2014年7月至2022年12月间北京协和医院行腹部薄层增强CT扫描的177例患者、200个PCLs病灶的基线临床和影像学资料,根据随访3年间PCLs病灶是否出现欧洲PCLs研究小组指南规定的手术征象,将其分为进展组和无进展组。以3∶1的比例将200个PCLs病灶随机分为训练集(150个)和内部测试集(50个);以2011年10月至2020年5月间南京大学医学院附属金陵医院行腹部薄层增强CT扫描的14例患者、15个PCLs病灶作为外部测试集。记录患者的临床及CT影像学特征,采用多种特征选择方法、多种机器学习模型进行组合,并基于10折交叉验证法筛选最优机器学习模型。绘制各模型的受试者工作特征曲线(ROC),计算曲线下面积(AUC),取AUC值最高的模型作为最优模型。在测试集上计算AUC值、灵敏度、特异度和准确度,评估模型的预测效能。使用置换重要性评估最优模型特征的重要程度。建立最优模型的校准曲线,采用Hosmer-Lemeshow检验评估模型的临床适用性。结果训练集和内部测试集的进展组与无进展组在胰腺炎史、病灶大小、主胰管管径、主胰管扩张、囊壁增厚、分隔存在、分隔增厚间的差异及内部测试集的两组在性别、病灶钙化和胰腺萎缩间的差异均有统计学意义(P值均<0.05)。外部测试集的进展组与无进展组在病灶大小和胰管扩张间的差异均有统计学意义(P值均<0.05)。基于F检验所选出的胰腺炎史、病灶大小、囊壁增厚、主胰管扩张、主胰管管径5个特征建立的支持向量机模型在交叉验证过程中取得了最高AUC值(0.899)。该模型在内部测试集中对PCLs进展预测的AUC值为0.909,灵敏度为82.4%,特异度为72.7%,准确度为76.0%,在外部测试集中分别为0.944、100%、77.8%和86.7%。�Objective To construct a machine-learning model for predicting the progression of pancreatic cystic lesions(PCLs)based on clinical and CT features,and to evaluate its predictive performance in internal/external testing cohorts.Methods Baseline clinical and radiological data of 200 PCLs in 177 patients undergoing abdominal thin slice enhanced CT examination at Peking Union Medical College Hospital from July 2014 to December 2022 were retrospectively collected.PCLs were divided into progressive and non-progressive groups according to whether the signs indicated for surgery by the guidelines of the European study group on PCLs were present during three-year follow-up.200 PCLs were randomly divided into training(150 PCLs)and internal testing cohorts(50 PCLs)at the ratio of 1∶3.15 PCLs in 14 patients at Jinling Affiliated Hospital of Medical School of Nanjing University from October 2011 to May 2020 were enrolled as external testing cohort.The clinical and CT radiological features were recorded.Multiple feature selection methods and machine-learning models were implemented and combined to identify the optimal machine-learning model based on the 10-fold cross-validation method.Receiver operating characteristics(ROC)curve was drawn and area under curve(AUC)was calculated.The model with the highest AUC was determined as the optimal model.The optimal model's predictive performance was evaluated on testing cohort by calculating AUC,sensitivity,specificity and accuracy.Permutation importance was used to assess the importance of optimal model features.Calibration curves of the optimal model were established to evaluate the model's clinical applicability by Hosmer-Lemeshow test.Results In training and internal testing cohorts,the progressive and non-progressive groups were significantly different on history of pancreatitis,lesions size,main pancreatic duct diameter and dilation,thick cyst wall,presence of septation and thick septation(all P value<0.05)In internal testing cohort,the two groups were significantly different on

关 键 词：胰腺囊性肿瘤 机器学习 疾病进展 计算机体层成像 

分 类 号：R735.9[医药卫生—肿瘤]