中文摘要：

目的基于中西医多模态信息并应用支持向量机技术（SVM）构建关于≥2级放射性肺损伤（RILI）早期预测模型。方法 2013年1月—2015年7月,共纳入接受放疗的肺癌患者191例,收集放疗前中医证候情况,放疗剂量学资料及其他一般临床信息。根据NCI-CTC3.0标准评价RILI级别,将放疗结束后发生≥2级RILI与否作为分类标签,随机选取119例病例作为建模组,剩余72例作为测试组。在SVM模型训练中,应用PCA法对于维度特征进行了降维处理,并应用交叉验证方法进行参数寻优,确立最佳参数,建立模型后通过验证集进行验证评价。结果在测试集建模中,5折交叉验证方法参数寻优确立了最佳参数,其中惩罚参数c为21.112,核函数参数g为0.009,初步建立模型,并通过验证集进行验证,验证结果示该SVM模型具有较强的预测效力,其敏感性为70.0%,特异性为75.2%,正确率73.8%,准确率为50.0%,F值为62.7%,AUC值0.72。结论基于SVM技术构建早期预测≥2级RILI具有一定的敏感性和特异性,提示该模型值得进一步验证并应用。

英文摘要：

Objective To develop and validate a model predicting radiation-induced lung injury（RILI）based on multi-modal information from Chinese and Western medicine in lung cancer patients receiving definitive thoracic radiation.Methods Between January 2013 and July 2015,a cohort of 191 patients receiving non-surgical treatment from Department of Radiation Oncology in author＇s hospital were included in this study.The information of TCM syndromes of patient and other clinical characteristics before radiotherapy were collected.National cancer institute common toxicity criteria version3.0was employed to evaluate the classification of RILI and grade≥2 toxicity served as the positive classification label.Patients were divided randomly into a study group（n=191）and a test group（n=72）.During model training,the PCA dimensionality reduction techniques were used to reduce the length of the feature vectors.K-fold Cross Validation（K-CV）was used to find the best parameter.Model predictive ability was evaluated in test model.Results During model training,the best parameter c and d were selected as 21.112 and 0.009 respectively by 5-fold CV method.The model developed in the model group was evaluated in the test group.The sensitivity,specificity,accuracy,precision,F score and AUC score were 70.0%,75.2%,73.8%,50.0%,62.7% and 0.71 respectively.Conclusion The SVM model which is constructed to predict RILI is a powerful and robust predictor and worth further testing.