基于SHAP方法的颅内动静脉畸形相关癫痫影像学特征分析

Radiomic feature analysis for intracranial arteriovenous malformation with epilepsy based on Shapley additive explanation method

目的通过机器学习算法和SHAP方法探讨与颅内动静脉畸形(AVM)患者出现症状性癫痫相关的关键影像学特征。方法回顾性纳入2022年1月至2023年1月北京医院(134例)、首都医科大学附属北京天坛医院(100例)、太原市中心医院(20例)、驻马店市中心医院(17例)4个中心的神经外科收治的271例颅内AVM患者, 其中首发症状为癫痫者31例(癫痫组), 按照1 ∶2的比例对年龄和性别进行倾向性评分匹配, 纳入对照组患者62例。基于CT血管成像资料比较两组患者病灶的体积、表面积、平均密度、球度、平坦度、伸长度、所在部位等传统影像学和形态学信息, 同时提取病灶的影像组学特征, 基于XGBoost建立机器学习模型, 通过受试者工作特征(ROC)曲线的曲线下面积评价模型预测颅内AVM患者是否伴发癫痫的性能, 并使用SHAP方法分析各影像学特征的重要性。结果癫痫组病灶位于颞叶5例(16.1%), 对照组病灶位于颞叶19例(30.6%), 两组的差异无统计学意义(χ^(2)=1.42, P=0.208)。两组患者病灶的体积、表面积、平均密度、球度、平坦度、伸长度的差异均无统计学意义(均P>0.05)。影像组学特征经过L1正则化降维, 最终纳入6项, 与传统影像学及形态学信息共同纳入XGBoost模型, 模型训练集和独立测试集的ROC曲线下面积分别为1.00和0.80。SHAP分析结果显示, 5项颅内AVM相关症状性癫痫的特征(包括病灶高伸长度、高平坦度、低平均密度、位于颞叶及低峰度)及其他3个影像组学特征对模型具有中等或较大贡献。结论伴发癫痫的颅内AVM患者的畸形团在传统影像学、形态学及影像组学特征上存在显著特点, 其中病灶的平坦度、伸长度、平均密度、位于颞叶及峰度为最重要的5个癫痫相关的影像学特征。

Objective To explore the key imaging features associated with the presence of epilepsy in patients with intracranial arteriovenous malformations(AVMs)using machine learning algorithms and Shapley additive explanation(SHAP)method.Methods A total of 271 patients with intracranial AVMs admitted to the Neurosurgery Departments of four centers including Beijing Hospital(134 patients),Beijing Tiantan Hospital of Capital Medical University(100 patients),Taiyuan Central Hospital(20 patients)and Zhumadian Central Hospital(17 patients)from January 2022 to January 2023 were retrospectively enrolled into this study.Thirty-one patients had the initial symptom of epileptic seizure(epilepsy group)and 62 patients(control group)were included in the proportion of 2∶1 by using the propensity score method.Traditional imaging and morphological features of lesions such as volume,surface area,mean density,roundness,flatness,elongation and location were compared between the two groups based on CT angiography data.While the radiomic features of the lesions were extracted,and a machine-learning model was built based on XGBoost.The performance of the model was evaluated by the area under the receiver operating characteristic curve(ROC)and the SHAP method was applied for assessing the importance of each feature.Results There were 5 cases(16.1%)of lesion located in the temporal lobe in the epilepsy group and 19 cases(30.6%)of lesion located in the temporal lobe in the control group,and the difference between the two groups was not statistically significant(χ2=1.42,P=0.208).There were no significant difference in volume,surface area,mean density,roundness,flatness,or elongation of the lesions between the two groups(all P>0.05).Radiomics features were downscaled by L1 regularization and 6 features were finally included,which were incorporated into the XGBoost model together with traditional risk factors and morphological features.The area under the ROC was 1.00 in training set and 0.80 in the independent test set.The results of the SHAP analysis showed that 5 features of epilepsy associated with intracranial AVMs,which included lesions with higher elongation length,higher flatness,lower mean density,being located in the temporal lobe and lower kurtosis and other 3 radiomics features had moderate or large contribution to the model.Conclusions Patients with intracranial AVMs associated with epilepsy have significant traditional imaging,morphological and radiomic features.Among them,flatness,elongation,mean density,location in the temporal lobe and kurtosis are the five most important features of epilepsy-relevant imaging features.