基于随机森林算法检出hs-cTnT项目小偏移的实时质控方法  

A real-time quality control method for detecting small shift of hs-cTnT based on random forest algorithm

作　　者：梁玉芳[1] 程华丽 王清涛[1,3] 王哲 冯祥 韩泽文 宋彪 周睿 LIANG Yufang;CHENG Huali;WANG Qingtao;WANG Zhe;FENG Xiang;HAN Zewen;SONG Biao;ZHOU Rui(Department of Laboratory Medicine,Beijing Chaoyang Hospital,Capital Medical University,Beijing 100020;Department of Laboratory Medicine,Aviation General Hospital,Beijing 100012;Beijing Center for Clinical Laboratories,Beijing 100020;Inner Mongolia Wesure Data Technology Co.,Ltd,Hohhot 010000,Inner Mongolia;Inner Mongolia University of Finance and Economics,Hohhot 010051,Inner Mongolia,China)

机构地区：[1]首都医科大学附属北京朝阳医院检验科,北京100020 [2]航空总医院检验科,北京100012 [3]北京市临床检验中心,北京100020 [4]内蒙古卫数数据科技有限公司大数据研究院,呼和浩特010000 [5]内蒙古财经大学,呼和浩特010051

出　　处：《临床检验杂志》2021年第12期950-955,共6页Chinese Journal of Clinical Laboratory Science

基　　金：北京市临床重点专科卓越项目(检验科)。

摘　　要：目的利用机器学习随机森林(random forest,RF)算法,建立一种准确识别检测小偏移的实时质控方法,并以浮动异常值之和的方法(moving sum of outlier,MovSO)作为参比方法,评价新算法效能。方法收集来自航空总医院实验室信息系统导出的2016年1月至2021年8月在罗氏化学发光E601设备检测的hs-cTnT项目检测结果,按照规定的数据清洗规则筛选出54243个结果作为无偏数据,人为引入10个不同大小的偏移,生成相应的有偏数据,每种偏移下用RF与MovSO两种算法进行实验。采用分类模型标准及临床指标对算法进行评价。结果RF算法在浮动窗口大小为10时,对10个小偏移均能检出,假阳性率(FPR)为4.0%~4.7%,受影响的患者样本数中位数(MNPed)在12以下。除了在±1 ng/L偏移时准确度为85%,其余8个偏移检出准确度均在90%以上;MovSO算法的最优浮动窗口大小为200,除了在1 ng/L时偏移无法检出,对其他9个偏移均可检出,FPR在3.5%~4.6%之间,MNPed均在100以上,仅在5 ng/L偏移时识别准确度方可达89%。RF算法总体显著优于MovSO,RF可准确识别hs-cTnT的测量小偏移。结论基于机器学习RF算法建立的质控方法可以改进类似hs-cTnT等临床对检测质量要求较高的项目的测量准确度,为实验室质控方案提供了新思路。Objective To establish a real-time quality control(QC)method for accurately detecting small shift by using the random forest(RF)algorithm of machine learning(ML),and evaluate the effectiveness of the new algorithm by taking the moving sum of outlier(MovSO)algorithm as a reference method.Methods The results of hs-cTnT detected on Roche E601 during January 2016 and August 2021 were collected through the laboratory information system of Aviation General Hospital.A total of 54243 results were selected as unbiased data in accordance with pre-defined data cleaning rules.Then,10 different shifts were artificially introduced to generate corresponding biased data.Next,RF and MovSO algorithms were tested for each shift.Finally,classification model criteria and clinical indexes were used to evaluate the two algorithms.Results When the block size of RF algorithm was defined as 10,all shifts could be detected,false positive rates(FPR)were between 4.0%and 4.7%,and all of MNPeds were less than 12.Except that the accuracy was 85%at±1 ng/L,the detection accuracy of the other 8 shifts was more than 90%.However,the optimal block size of MovSO algorithm was 200.Except that the shift could not be detected at 1ng/L,the other 9 shifts could be detected.The FPRs were between 3.5%and 4.6%,and all of MNPeds were more than 100.The accuracy could reach 89%only at the largest shift of 5ng/L.In general,RF algorithm was significantly better than MovSO,and RF could accurately identify the analytical small shift of hs-cTnT.Conclusion The QC method based on ML algorithm can effectively improve the detection accuracy of test items with high quality requirements such as hs-cTnT,and provide a new idea for laboratory QC.

关 键 词：机器学习 随机森林 小偏移 实时质控 实验室数据 

分 类 号：R446[医药卫生—诊断学]