临床-影像组学联合模型预测自发性脑出血后早期血肿扩大的研究

Clinical-radiomics combined model in prediction of early hematoma expansion after spontaneous intracerebral hemorrhage

目的:探讨自发性脑出血(sICH)患者出血后早期血肿扩大(HE)的危险因素,构建预测sICH后早期HE的临床-影像组学联合模型并评估其预测效能。方法:选择重庆医科大学附属第二医院放射科自2014年4月至2020年9月行CT平扫的sICH患者339例,根据是否发生HE将患者分成HE组和非HE组(HE定义为24 h内复查CT图像中血肿体积较首次CT平扫图像中增加>33%或6 mL)。比较非HE组和HE组患者的临床资料,采用多因素Logistic回归分析筛选HE发生的危险因素。提取患者首次CT平扫图像中脑血肿兴趣区的影像组学特征,采用LASSO回归模型和10折交叉验证法筛选最优影像组学特征并计算影像组学评分(R-score),利用HE发生的危险因素、R-score分别构建临床模型、R-score和临床-影像组学联合模型,采用受试者工作特征曲线(ROC)分析各模型对HE的预测效能,将最佳模型可视化为列线图并绘制校准曲线评估列线图模型预测的准确性。结果:与非HE组比较,HE组患者中发病至首次CT检查时间较短、糖尿病者所占比例较高、血小板计数较低、格拉斯哥昏迷量表(GCS)评分较低、首次CT图像中血肿体积较大,差异均有统计学意义(P<0.05)。多因素Logistic回归分析显示首次CT图像中血肿体积(OR=1.015,95%CI:1.000~1.030,P=0.046)、GCS评分(OR=0.914,95%CI:0.839~0.995,P=0.039)、发病至首次CT检查时间(OR=0.855,95%CI:0.741~0.987,P=0.032)和糖尿病(OR=0.522,95%CI:0.311~0.875,P=0.014)是发生HE的独立危险因素。LASSO回归模型和10折交叉验证法最终筛选出20个最优影像组学特征。临床模型、R-score和临床-影像组学联合模型预测HE的曲线下面积分别为0.650、0.860和0.870。校准曲线显示临床-影像组学联合模型对早期HE的预测概率与实际发生概率具有较好的一致性。结论:临床-影像组学联合模型能够有效地预测早期HE且具有良好校准度,有助于临床个体化评估sICH患者发生早期HE的风险。

Objective To explore the risk factors for early hematoma expansion(HE)in patients with spontaneous intracerebral hemorrhage(sICH),and construct a clinical-radiomics combined model to predict HE after sICH.Methods From April 2014 to September 2020,339 patients with sICH who underwent plain CT scans in Radiology Department of our hospital were recruited.Patients were divided into HE group and non-HE group according to whether HE occurred(HE was defined as an increase in hematoma volume>33%or 6 mL on the follow-up CT within 24 h).The clinical data of non-HE group and HE group were compared,and multivariate Logistic regression analysis was used to detect independent risk factors for HE.The radiomics features were extracted from the regions of interest of the hematoma in the first CT scan images;the optimal radiomics features were selected using least absolute shrinkage and selection operator(LASSO)regression model and 10-fold cross-validation method,and then,the radiomics scores(R-score)were calculated;the risk factors for HE(clinical data)and R-score(radiomics data)were used to construct the clinical model,R-score model,and clinical-radiomics combined model;receiver operating characteristic(ROC)curve was performed to evaluate the prediction performance of clinical model,R-score model,and clinical-radiomics combined model;the best model was visualized as a nomogram and a calibration curve was drawn to evaluate the prediction accuracy of this model.Results As compared with patients in the non-HE group,patients in the HE group had shorter time from sICH onset to first CT,higher percentage of patients with diabetes,lower platelet count,lower Glasgow Coma Scale(GCS)scores,and larger baseline hematoma volume in CT image,with significant differences(P<0.05).Multivariate Logistic regression analysis showed that baseline hematoma volume(OR=1.015,95%CI:1.000-1.030,P=0.046),GCS scores(OR=0.914,95%CI:0.839-0.995,P=0.039),time from sICH onset to first CT(OR=0.855,95%CI:0.741-0.987,P=0.032),and diabetes(OR=0.522,95%CI:0.311-0.875,P=0.014)were independent risk factors for HE.By using LASSO regression and 10-fold cross-validation method,20 optimal radiomics features were finally selected.The area under ROC curve of clinical model,R-score model,and clinical-radiomics combined model were 0.650,0.860,and 0.870,respectively.The calibration curve showed that the prediction accuracy of clinical-radiomics combined model in early HE had good consistency with the actual occurrence probability.Conclusion The clinical-radiomics combined model could effectively predict early HE with good calibration,which is helpful in individualized clinical assessment of risk of early HE in SICH patients.