基于SHAP值特征选择的γ通过率分类预测及解释  

Gamma pass rate classification prediction and interpretation based on SHAP value feature selection

作　　者：陈路桥 倪千喜 庞金猛 谭剑锋 周新 骆龙军 曾德高 曹锦佳[2] Chen Luqiao;Ni Qianxi;Pang Jinmeng;Tan Jianfeng;Zhou Xin;Luo Longjun;Zeng Degao;Cao Jinjia(Department of Radiation Oncology,Hunan Cancer Hospital/the Affiliated Cancer Hospital of Xiangya School of Medicine,Central South University,Changsha 410013,China;School of Nuclear Science and Technology,University of South China,Hengyang 421001,China)

机构地区：[1]湖南省肿瘤医院/中南大学湘雅医学院附属肿瘤医院放疗科,长沙410013 [2]南华大学核科学技术学院,衡阳421001

出　　处：《中华放射肿瘤学杂志》2023年第10期914-919,共6页Chinese Journal of Radiation Oncology

基　　金：湖南省自然科学基金面上项目(2023JJ30373);湖南省科技创新计划资助项目(2021SK51116);湖南省卫生健康委科研计划项目(20201364);湖南省卫生健康委适宜技术推广项目(202218015767)。

摘　　要：目的:探索SHAP值结合极端梯度提升树(XGBoost)算法的特征选择技术来构建调强放疗γ通过率预测模型的可行性和有效性,并给出相应的模型解释。方法:回顾性分析2020年11月至2021年11月在湖南省肿瘤医院接受盆腔固定野调强放射治疗的196例肿瘤患者采用基于模体测量方式的调强放疗计划的剂量验证结果,γ通过率标准为3%/2 mm、10%剂量阈值。提取基于剂量文件的影像组学特征并使用SHAP值结合XGBoost算法进行特征筛选后构建预测模型。分别选取特征数量为50、80、110、140个,构建四种机器学习分类模型,计算曲线下面积(AUC)值、召回率及F1分数评估预测模型的分类性能。结果:基于SHAP值特征选择的110个特征构建的预测模型AUC值为0.81,召回率达到0.93,F1分数为0.82,均优于其他三个模型。结论:针对盆腔肿瘤调强放疗计划,可以采用SHAP值与XGBoost算法结合以选择用于预测的最佳影像组学特征子集来构建γ通过率的预测模型,并能通过SHAP值给出模型输出解释,可能在理解依赖机器学习模型所做的预测方面提供价值。Objective To explore the feasibility and validity of constructing an intensitymodulated radiotherapy gamma pass rate prediction model after combining the SHAP values with the extreme gradient boosting tree(XGBoost)algorithm feature selection technique,and to deliver corresponding model interpretation.Methods The dose validation results of 196 patients with pelvic tumors receiving fixed-field intensity-modulated radiotherapy using modality-based measurements with a gamma pass rate criterion of 3%/2 mm and 10%dose threshold in Hunan Provincial Tumor Hospital from November 2020 to November 2021 were retrospectively analyzed.Prediction models were constructed by extracting radiomic features based on dose files and using SHAP values combined with the XGBoost algorithm for feature filtering.Four machine learning classification models were constructed when the number of features was 50,80,110 and 140,respectively.The area under the receiver operating characteristic curve(AUC),recall rate and F1 score were calculated to assess the classification performance of the prediction models.Results The AUC of prediction model constructed with 110 features selected based on the SHAP-valued features was 0.81,the recall rate was 0.93 and the F1 score was 0.82,which were all better than the other 3 models.Conclusion For intensity-modulated radiotherapy of pelvic tumor,SHAP values can be used in combination with the XGBoost algorithm to select the optimal subset of radiomic features to construct predictive models of gamma pass rates,and deliver an interpretation of the model output by SHAP values,which may provide value in understanding the prediction by machine learning-dependent models.

关 键 词：机器学习 调强放射疗法 特征选择 模型解释 γ通过率 

分 类 号：R730.55[医药卫生—肿瘤]