无监督学习在临床实验室自动审核规则制定中的实践价值

The clinical value of unsupervised learning in the establishment of clinical laboratory autoverification rules

目的探索自编码机算法对临床实验室自动审核规则制定的临床价值。方法通过机器学习算法中的自编码机算法对四川省达州市中心医院检验科过去3年已经人工审核的涉及血清总蛋白(TP)、血清清蛋白(ALB)以及免疫球蛋白的结果进行无监督学习,评估选择合理模型的部分指标,以及对模型分组后的数据进行分布分析,进而探讨模型的合理性。结果通过增加合适的范围内隐层和隐层内神经元的数量,以及采用dropout技术会对最终模型的MSE带来积极的效果;本文模型认为TP介于44.59～94.56g/L之间,ALB位于27.14～52.88g/L之间,A/G值位于0.65～4.86之间,IgA位于0.01～18.74g/L之间,IgG位于1.34～36.81g/L之间,及IgM位于0.001～10.715g/L之间的结果可以无需关注,该判断规则与部分实验室已公布的自动审核规则极为接近,本试验数据显示将近90%判定为无需关注,进而大大提高检验人员对极少数异常数据的关注度。结论无监督学习中的自编码机算法能够在实验室自动审核规则制定中产生积极的价值,但需要进一步对模型产生的规则进行深入解读和分析。

Objective To explore the clinical value of autoencoder machine algorithm in the establishment of clinical laboratory autoverification rules.Methods Experimental models were generated by deep autoencoder algorithm.Dataset confirmed by manual review was received from clinical laboratory of Dazhou Central Hospital,Sichuan Province.The items observed included serum total protein（TP）,albumin（ALB）and immunoglobulin.The distribution characteristics of grouped data were analyzed in order to investigate the validity of the model.Results By increasing the number of hidden layers and neurons in the appropriate range,and using dropout technology,the MSE of final model could be positively affected.The rules of the final model,TP=44.59-94.56 g/L,ALB=27.14-52.88 g/L,A/G=0.65-4.86,IgA=0.01-18.74 g/L,IgG=1.34-36.81 g/L and IgM=0.001-10.715 g/L,are very close to the atuoverification rules published by some laboratories,and can effectively determine the extreme abnormal data of TP,ALB,A/G,and high level of Ig.This model can determine 90%of the data without concern,and thus greatly improve the attention of inspectors to a small number of abnormal data.Conclusion The autoencoder algorithm can generate positive value in the formulation of autoverification rules in laboratory,but further interpretation of the model as well as in-depth analysis is needed.