基于机器学习的MRI放射组学急性缺血性脑卒中复发预测模型研究

Research on the Predictive Model of MRI Radiomics Acute Ischemic Stroke Recurrence Based on Machine Learning

目的探讨基于MRI放射组学的急性缺血性脑卒中(AIS)患者2年内复发风险预测模型研究。方法回顾性搜集148例AIS患者的MRI和临床资料。根据患者出院后2年内复发情况分为复发组和非复发组。为了更好的验证模型性能,按7∶3的比例将样本随机分为训练集和测试集。由两名放射科医师使用MaZda软件包在颅脑扩散加权成像(DWI)MRI序列上标记病变,从图像中提取放射组学特征,并使用LASSO Logistic回归模型对放射组学特征进行降维处理。采用20种常见机器学习算法构建基于MRI放射组学特征结合临床特征的预测模型,并根据最优模型参数构建单独的MRI放射组学特征模型和临床特征模型后加以比较说明。采用敏感度、特异度和受试者工作特征曲线下面积等指标来比较模型的预测性能。结果从DWI图像中的300个放射组学特征中LASSO-Logistic最终提取了14个特征。放射组学标签值在未复发和复发患者的分布差异显著(P<0.05)。基于MRI和临床结合的模型中,XGBoost模型在测试集和训练集中均获得了较好的预测精度,且较单纯模型性能较佳。SHAP分析表明Radscore对模型的影响最大,且对AIS复发概率是正贡献;其次为高密度脂蛋白,是负贡献;然后是甘油三酯和低密度脂蛋白,为正贡献;最后是年龄和脂蛋白为正贡献;此外发现总胆固醇和性别对模型影响不大。结论基于从MRI放射组学和临床指标相结合的XGBoost模型在预测AIS术后2年内的复发性能方面表现最佳。基于MRI放射组学特征与临床数据的结合提高了模型的预测性能。

Objective This study aimed to explore a prediction model for the risk of recurrent AIS within 2 years based on MRI radiomics.Methods We retrospectively collected MRI and clinical data from 148 patients diagnosed with AIS at the affiliated hospital of Xuzhou Medical University between August 2018 and July 2019.The patients were divided into a recurrent group and a non-recurrent group based on recurrence within 2 years after discharge.To validate the model performance,the samples were randomly divided into training and testing sets with a 73 ratio.Two radiologists marked the lesions on diffusion-weighted imaging(DWI)MRI sequences using MaZda software,extracted radiomic features from the images,and reduced the dimensionality of the radiomic features using LASSO-Logistic regression.Twenty common machine learning algorithms were used to construct prediction models combining MRI radiomic features with clinical features.The individual MRI radiomic feature model and clinical feature model were compared based on the optimal model parameters.Model performance was evaluated using metrics such as sensitivity,specificity,and area under the receiver operating characteristic curve(AUC).Results Fourteen features were finally extracted from the 300 radiomic features using LASSO-Logistic from DWI images.The distribution of radiomic signature values differed significantly between non-recurrent and recurrent patients(P<0.05).The XGBoost model in the combined MRI and clinical model showed good predictive accuracy in both the testing and training sets,outperforming the single-feature models.SHAP analysis revealed that Radscore had the greatest impact on the model and made positive contributions to the AIS recurrence probability,followed by HDL-C with negative contributions.TG and LDL-C made positive contributions,while age and LP(a)made positive contributions.Additionally,TC and gender showed minimal impact on the model.Conclusion The XGBoost model based on combined MRI radiomic features and clinical indicators performed best in predicting recurrent AIS within 2 years.Combining MRI radiomic features with clinical data improved the predictive performance of the model.