泌尿系结石CT征象与手术方式的相关性：NLP研究

目的:以自然语言处理(NLP)方法研究泌尿系结石CT报告征象与临床手术方式的相关性.方法:回顾性选取HIS中因泌尿系结石行CT检查、并在本院行外科处理的住院患者资料,包括CT报告、出院小结中记录的诊疗经过,形成记录文档.使用不同NLP模型研究CT报告中结石的影像特征与手术方式的相关性.结果:共纳入371例连续患者的资料用于分析,其中行经皮肾镜碎石取石术(PNL) 142例,输尿镜碎石术(URL) 190例,输尿管软镜碎石术(IRS) 39例.NLP提取“结石”、“炎症”、“输尿管形态”为关键CT特征,十折交叉验证结果显示,经优化后的NLP对结石手术(URL和PNL)分类符合率为0.730,ROC曲线下面积为0.741.结论:NLP用于CT结石特征分析是可行的,可为临床决策提供一定的帮助.

Correlation between CT signs and surgical procedures of urinary calculus:optimization of natural language processing research

Objective:To investigate the correlation between CT report signs and surgical procedures of urinary calculus by naturallanguage processing (NLP) method.Methods:Documents were retrospectively archived from HIS for the inpatients,who had CT examination for urinary calculus and ultimately underwent surgical procedures.NLP model was constructed and tested in the following steps:word segmentation,vector construction of word bags,word frequency construction and feature vector construction.Decision tree based on ngrams and RPART algorithm was used for correlation analysis between CT features and surgical methods.Results:Totally 371 consecutive data of patients were recruited for NLP analysis,including cases of 142 percutaneous nephrolithotripsy (PNL),190 ureteroscopic lithotripsy (URL) and 39 retrograde intrarenal surgery (RIRS).According to NLP,there was correlation between the CT features of "calculus","inflammation",and "morphology of ureter" to the choice of operation method.10-fold cross validation demonstrated that the precision for PNL and URL was 0.730,and area under ROC curve was 0.741.Conclusion:It is feasible to use the natural language processing in the decision-making of operation treatment for urinary calculus.