Abstract: To develop a predictive model for the occurrence of parastomal hernia (PSH) after colostomy and validate it. Methods    A total of 131 rectal cancer patients who underwent permanent colostomy in our hospital from January 2016 to December 2020 were retrospectively analyzed and their preoperative clinical information and abdominal CT were collected. Patients were divided into PSH group（n=43） and nonPSH group （n=88）according to whether PSH occurred. Then, the PSH and nonPSH groups are divided into training and validation sets in the ratio of 7∶3, respectively. The preoperative abdominal CT images of the 3rd lumbar vertebral level were segmented to extract the radionics features, and at the same time, the preoperative clinical indicators were collected and screened. Three algorithms, support vector machine (SVM), decision tree (DT), and random forest (RF), were used to construct a prediction model by incorporating the screened clinical indicators and radionics features. The accuracy, sensitivity, specificity, and area under the ROC curve (AUC) of the models were calculated to evaluate the predictive efficacy of different models. Results    In terms of clinical indicators, the total serum protein and BMI in the PSH group were significantly higher than those in the no-PSH group（P<0.05）. A total of 6 non-zero coefficient features were selected from the 107 extracted radionics features through LASSO regression. The AUCs of the prediction models built by SVM, DT, and RF were 0.820, 0.854, and 0.790 respectively in the training set; they were 0.804, 0.762, and 0.732 respectively in the validation set. Conclusion    Using patients' preoperative clinical examination data and CT images can help identify high-risk groups for PSH and provide a reference for the prevention and clinical personalized diagnosis and treatment.

Key words: parastomal hernia, radiomics, machine learning, prediction model

摘要： 目的    建立结肠造口术后造口旁疝（PSH）发生预测模型，并对模型进行验证。方法    回顾性分析2016年1月至2020年10月陆军军医大学第二附属医院普通外科收治的行结肠末端造口的131例乙状结肠癌或直肠癌病人的临床资料。按是否发生PSH将病人分为PSH组（43例）和无PSH组（88例）；再按7∶3的比例对PSH组和无PSH组分别进行训练集和验证集的随机划分。将术前腹部CT图像第3腰椎轴向截面腹壁感兴趣区域分割后提取影像组学特征，同时收集术前临床指标并筛选。利用支持向量机（SVM） 、决策树（DT）和随机森林（RF）3种算法，纳入筛选的临床指标和影像组学特征构建预测模型。计算准确率、敏感度、特异度和受试者工作特征曲线下面积（AUC）来评价不同模型预测效能。结果    在临床指标方面，PSH组病人的血清总蛋白和BMI高于无PSH组，差异有统计学意义（P<0.05）。在影像特征方面，通过LASSO回归降维筛选得到6个非零系数特征。SVM、DT和RF构建预测模型的AUC在训练集中分别为0.820、0.854、0.790，在验证集中分别为0.804、0.762、0.732。结论    根据病人术前临床指标及CT影像结合机器学习构建预测模型，有助于识别PSH高危人群，可为PSH预防和临床个体化诊疗提供参考。

关键词: 造口旁疝, 影像组学, 机器学习, 预测模型