基于颅内斑块影像组学联合传统标志物预测缺血性脑卒中复发风险

Prediction of the risk of recurrent ischemic stroke based on intracranial plaque radiomics with traditional biomarkers

目的对症状性大脑中动脉斑块进行影像组学特征提取,建立Cox回归模型来预测2年随访期间非心源性前循环缺血性脑卒中的复发风险,并评估模型性能。材料与方法回顾性分析2019年1月至2020年1月完成高分辨磁共振成像(high-resolution magnetic resonance imaging,HRMRI)的首发急性前循环大脑中动脉>50%狭窄的缺血性脑卒中患者,采集其基线数据及2年随访信息,终点事件定义为随访期间非心源性的缺血性脑卒中复发事件,最终纳入82例合格受试者。利用3D-slicer软件连续进行手动分割及Radiomics插件进行斑块纹理提取,在最小绝对收缩和选择算子(least absolute shrinkage and selection operator,LASSO)算法回归分析的最优超参数指导下建立Cox回归预测模型,并进一步计算C指数及绘制校准曲线,以进一步评估模型性能。结果共19例患者在2年内随访期间出现终点事件,平均随访时间19.9个月,缺血性脑卒中复发率为13.9/100人年。本文根据最优超参数(logλmin=0.0886)所确定的前4个主成分(二阶纹理特征、斑块内出血、熵及低密度脂蛋白胆固醇),将其纳入多因素Cox回归分析并校正基线协变量,发现斑块二阶纹理特征(即灰度共生矩阵)对模型贡献价值最大,即每增加1个风险积分,2年内缺血性脑卒中复发风险升高约5倍[校正风险比(adjusted hazard ratio,aHR):5.379,95%置信区间(confidence interval,CI):1.716~16.859,P=0.004,权重=40.23%];然而,斑块内出血(aHR:2.226,95%CI:0.821~6.040,P=0.116,权重=20.86%)、熵(aHR:1.324,95%CI:0.769~2.278,P=0.311,权重=16.13%)及低密度脂蛋白胆固醇(aHR:1.485,95%CI:0.877~2.516,P=0.142,权重=22.78%)对终点事件预测呈相对较弱的正相关趋势。进一步在模型评价方面,文中建立的Cox风险预测模型展现了优秀的模型性能(区分度C指数=0.8296;贝叶斯校正的曲线显示贴合45°对角线)。结论本研究建立的脑卒中复发的风险预测模型有助于筛选高复发风险的缺血性脑卒中患者,其中灰度共生矩阵对于复发风险的预测价值最大,这可能为未来精准治疗及开展大规模前瞻临床研究提供重要证据。

Objective:To develop Cox proportional hazards regression model for prediction of the recurrence risk of non-cardiogenic anterior circulation ischemic stroke during 2-year follow-up based on radiomic approach by extracting texture features from a symptomatic middle cerebral artery(MCA)plaque,and to further evaluate the developed model performance.Materials and Methods:In our retrospective study from January 2019 to January 2020,a total of 82 eligible patients with first-ever ischemic stroke and middle cerebral artery>50%luminal stenosis underwent baseline intracranial high-resolution magnetic resonance imaging(HRMRI)followed up for 2 year when recurrent non-cardiogenic ischemic stroke in the territory of MCA served as an endpoint event were finally enrolled in current analyses.HRMRI-based radiomic features were manually extracted from an index MCA plaque using 3D-Slicer software package.Multivariable Cox regression analysis was used to develop the predicting model where multi-dimensional parameters were selected by LASSO(least absolute shrinkage and selection operator)regression analysis,for which the performance was further assessed with respect to its calibration,discrimination.Results:Of which,19 cases with endpoint events occurred during the 2-year follow-up period with 13.9 per 100 person-years of the recurrence rate of ischemic stroke.Multivariable Cox regression included top 4 parameters with nonzero coefficients defined by logλmin of LASSO regression(i.e.,second-order texture feature,plaque hemorrhage,entropy and low density lipoprotein cholesterol).In the prediction model adjusting for baseline covariants,the gray level co-occurrence matrix was found to be the major contributor to the event endpoint[adjusted hazard ratio(aHR):5.379,95%confidence interval(CI):1.716-16.859,P=0.004,weight=40.23%].However,plaque hemorrhage(aHR:2.226,95%CI:0.821-6.040,P=0.116,weight=20.86%),entropy(aHR:1.324,95%CI:0.769-2.278,P=0.311,weight=16.13%)and low density lipoprotein cholesterol(aHR:1.485,95%CI:0.877-2.516,P=0.142,weight=22.78%)just showed a trend towards significance.Additionally,the developed prediction model showed a good discrimination with a C-index of 0.8296 and good calibration.Conclusions:The findings suggest that our developed prediction model can target a potential sub-population at high risk of recurrent ischemic stroke in which gray level co-occurrence matrix may account for the major contributing,although this must be confirmed in future.