基于脑胶质瘤（BG）患者术前增强MRI成像的Delta影像组学特征和临床基线资料构建预测模型，评估该模型对BG复发及生存期的预测价值。

回顾性分析湖北医药学院附属襄阳市第一人民医院肿瘤科自2015年8月至2023年4月收治的220例BG患者的临床资料，根据患者生存状态将其分为正常组（n=125）和复发组（n=95），以7∶3的比例建立训练集（155例，正常组88例和复发组67例）和测试集（65例，正常组37例和复发组28例）。分别从瘤周水肿区域（3~12 mm）及肿瘤内提取Delta影像组学特征，采用LASSO算法结合10折交叉验证生成瘤周放射组学评分（Radscore1）和瘤内放射组学评分（Radscore2）。比较训练集和测试集中正常组和复发组患者的临床基线资料及影像学特征差异。基于训练集数据，采用多因素Logistic回归构建3个术后复发预测模型：临床模型、影像模型和组合模型，采用DeLong检验比较各模型间受试者工作特征（ROC）曲线下面积（AUC）的差异，运用决策曲线分析评估各模型的临床净获益，使用XGboost机器学习算法对各模型的预测结果进行验证，并于测试集中进一步验证模型的预测效能。采用Kaplan-Meier法绘制术后总生存期（OS）曲线，组间比较采用Log-rank检验，计算中位生存期及其95%CI，并构建生存列线图评估其对BG患者术后生存的预测价值。

训练集中，2组患者的性别、异柠檬酸脱氢酶-1（IDH1）突变、1p19q杂合缺失、MRI强化方式、Radscore1及Radscore2比较，差异有统计学意义（P<0.05）。测试集中上述资料比较结果与训练集一致。多因素Logistic回归分析显示性别、IDH1突变状态、1p19q杂合缺失、Radscore1及Radscore2是BG复发的独立影响因素（P<0.05）。基于上述因素构建3个预测模型：临床模型、影像模型及组合模型，ROC曲线分析显示，组合模型的预测性能（AUC=0.890，95%CI：0.830~0.935）显著优于临床模型（AUC=0.748，95%CI：0.672~0.814）和影像模型（AUC=0.832，95%CI：0.763~0.887），差异均有统计学意义（P<0.05）。XGboost算法验证了组合模型的预测结果，决策曲线显示组合模型的临床净获益更高。该结果在测试集中得到验证，且最终基于组合模型的列线图简化了BG复发的预测过程，在测试集中敏感度0.857，特异度0.756。Kaplan-Meier生存曲线分析显示，性别、1p19q杂合缺失、Radscore1、Radscore2各亚组间OS比较，差异均有统计学意义（P<0.05）。基于这些因素建立的生存列线图预测BG生存期的AUC为0.859（95%CI：0.802~0.933），敏感度和特异度分别为0.872和0.705。

基于术前增强MRI的Delta影像组学预测模型可提高BG患者术后复发及生存期的预测准确性，为术后个体化管理和临床决策提供依据。

To construct a predictive model based on the Delta radiomics and clinical baseline data of preoperative contrast-enhanced MRI imaging, and evaluate its predictive value for recurrence and survival in patients with brain glioma (BG).

A retrospective analysis was conducted on clinical data of 220 BG patients admitted to Oncology Department of Xiangyang No.1 People’s Hospital, Hubei University of Medicine from August 2015 to April 2023. Patients were divided into a normal group (n=125) and a recurrence group (n=95) according to their survival status. The cohort was split into a training set (155 cases, 88 normal cases and 67 recurrence cases) and a testing set (65 cases, 37 normal cases and 28 recurrence cases) at a ratio of 7∶3. Delta radiomics features were extracted from the peritumoral edema area (3-12 mm) and within the tumor, respectively, and the least absolute shrinkage and selection operator (LASSO) algorithm combined with ten-fold cross-validation was used to generate the peritumoral radiomics score (Radscore1) and intratumoral radiomics score (Radscore2). The clinical baseline data and imaging characteristics of normal and recurrent patients in the training and testing sets were compared. Based on the training set data, 3 postoperative recurrence prediction models were constructed using multiple Logistic regression: clinical model, imaging model, and combination model. DeLong test was used to compare the differences in area under the ROC curve (AUC) between the models, and decision curve analysis was used to evaluate the clinical net benefit of each model. XGboost machine learning algorithm was used to verify the prediction results of each model, and the predictive performance of the models was further validated in the testing set. Kaplan-Meier method was used to plot the overall survival (OS) curve after surgery, and the Log-rank test was used for inter group comparison. The median OS and 95%CI were calculated, and a survival column chart was constructed to evaluate its predictive value for postoperative survival of BG patients.

In the training set, there were statistically significant differences in gender, isocitrate dehydrogenase-1 (IDH1) mutation, 1p19q heterozygous deletion, MRI enhancement mode, and Radscore1 and Radscore2 between the two groups of patients (P<0.05). The comparison results of the above data in the testing set are consistent with those in the training set. Multivariate Logistic regression analysis showed that gender, IDH1 mutation status, 1p19q heterozygous deletion, Radscore1 and Radscore2 were independent influencing factors for BG recurrence (P<0.05). Based on the above factors, three prediction models were constructed: clinical model, imaging model, and combination model. ROC curve analysis showed that the predictive performance of the combination model (AUC=0.890, 95%CI: 0.830-0.935) was significantly better than that of the clinical model (AUC=0.748, 95%CI: 0.672-0.814) and the imaging model (AUC=0.832, 95%CI: 0.763-0.887), and the differences were statistically significant (P<0.05). The XGboost algorithm validated the prediction results of the combination model, and the decision curve showed that the clinical net benefit of the combination model was higher. This result was validated in the testing set, and ultimately the column chart based on the combination model simplified the prediction process of BG recurrence, with a sensitivity of 0.857 and a specificity of 0.756 in the testing set. Kaplan-Meier survival curve analysis showed that there were statistically significant differences in OS between subgroups based on gender, 1p19q heterozygous deletion, Radscore1, and Radscore2 (P<0.05). The AUC for predicting BG survival based on the survival column chart established based on these factors is 0.859 (95%CI: 0.802-0.933), with sensitivity and specificity of 0.872 and 0.705, respectively.

The predictive model based on Delta radiomics from preoperative contrast-enhanced MRI can improve the prediction of postoperative recurrence and survival in BG patients, providing a basis for individualized management and clinical decision-making after surgery.