基于CT影像组学预测高血压性脑出血患者的临床结局

Prediction of clinical outcome of patients with hypertensive intracerebral hemorrhage based on CT radiomics

目的构建并验证一项列线图模型预测高血压性脑出血(HICH)患者30 d临床结局情况。方法采用双中心回顾性队列研究,按时间顺序将中心1的患者(n=431例)分为训练队列(n=305例)和内部验证队列(n=126例),中心2的患者作为外部验证队列(n=108例)。根据改良Rankin评分(mRS)对患者30 d的临床结局进行评估,mRS≤3为预后良好,mRS>3为预后不良。临床病史、实验室检查和CT平扫征象应用随机森林(RF)筛选出重要因素,重要因素纳入多因素Logistic回归分析,找出独立危险因素来建构临床模型。提取患者平扫CT影像组学的特征,利用最小绝对收缩和选择算子(LASSO)回归筛选出最优特征集,并对特征集进行影像组学评分(Radscore),从而构建影像组学模型。独立危险因素联合Radscore共同建立临床影像组学列线图模型。结果列线图模型由血肿部位、血肿体积、破入脑室、中线移位、GCS评分及Radscore预测因素构成,其在训练队列曲线下面积(AUC)为0.89,在内部和外部验证队列AUC分别为0.86和0.82。校准曲线分析显示,列线图在训练组队列中P=0.944,在内部验证队列中P=0.540,在外部验证队列中P=0.171。决策曲线分析(DCA)表明,列线图的阈值概率在(0.10~0.90)有较高的临床适用价值,最佳截断概率为0.33。结论列线图模型能有效的预测HICH患者30 d的临床结局情况,有一定临床诊疗指导价值。

Objective To construct and validate a nomogram model for predicting 30 days clinical outcomes of patients with hypertensive intracerebral hemorrhage(HICH).Methods In a two-center retrospective cohort study,patients in center 1(n=431 cases)were divided chronologically into a training cohort(n=305 cases)and an internal validation cohort(n=126 cases),and patients in center 2 formed an external validation cohort(n=108 cases).Clinical outcomes at 30 days were evaluated according to the modified Rankin score(mRS),with mRS≤3 indicating a good prognosis and mRS>3 indicating a poor prognosis.Clinical history,laboratory examination and noncontrast computed tomography scan signs were selected by random forest(RF),and important factors were included in multivariate Logistic regression analysis to identify independent risk factors to construct clinical models.The features of the patients’noncontrast computed tomography images were extracted,the optimal selection was selected by LASSO regression,and Radscore was performed on the feature set to construct the radiomics model.Independent risk factors and Radscore were combined to establish clinical radiomic nomogram model.Results The nomogram model is composed of hematoma location,hematoma volume,ventricular rupture,midline shift,GCS score and Radscore,and its AUC is 0.89 in the training cohort,and 0.86 and 0.82 in the internal and external verification cohort,respectively.The calibration curve analysis shows that the nomogram is in the training cohort P=0.944,in the internal verification cohort P=0.540 and in the external verification cohort P=0.171.DCA analysis shows that the threshold probability of nomogram is(0.10-0.90),which has high clinical application value,and the optimal truncation probability is 0.33.Conclusion The nomogram model can effectively predict the 30-day clinical outcome of patients with HICH,so it has a certain guiding value in clinical diagnosis and treatment.