基于智能数据和机器学习的尿液检验结果解释性报告

Interpreting report of urinalysis based on intelligent data and machine learning

目的:建立一种基于人工智能的尿液检验结果解释性报告系统。方法:收集2008—2018年浙江大学医学院附属第一医院患者2899917份、体检710971份尿检数据,统计每个项目不同结果的频数分布建立大人群分布,再根据数据分布、项目重要性和结果异常程度,建立每个样本的健康指数和各项目的异常等级。收集糖尿病、尿路感染、肾小球肾炎、肾病综合征等疾病数据,按性别、年龄匹配同数量的健康对照组。基于AdaBoost算法的集成学习器建立模型并评估算法性能。用JAVA开发数据展示软件。用199份异常尿液检验结果,人工验证模型的准确性。结果:每份报告分为正常、异常、疾病、危重4个等级;单个项目结果判断为正常、轻度、中度、重度、极度5个等级并提供大数据的人群分布;基于AdaBoost机器学习模型运用于7种疾病的训练准确度(≥88.3%)、真阳性率(≥80.0%)、曲线下面积(≥0.954)。开发的JAVA软件展示上述结果,并包括病历和结果、历史结果、个性化建议、异常项目科普、在大人群数据中的位置等内容。异常尿液结果可能的疾病相似度,人工验证机器学习模型的准确率为82.41%(164/199)。结论:本研究建立了智能的结果解释性报告系统,能区分报告异常程度,具有较高疾病预测准确性,可提供个性化的临床决策信息。

Objective:To establish an interpretive reporting system for urinalysis based on artificial intelligence(AI).Methods:Urine tests were collected from the First Affiliated Hospital,College of Medicine,Zhejiang University from 2008 to 2018,including 2899917 patient tests and 710971 physical check-up tests.Then we set up a large population distribution with the frequency of different results of each item and established a health index of each sample and an abnormal level of each item according to data distribution,importance and degree of abnormality.We collected data of seven diseases,such as diabetes mellitus,urinary tract infection,glomerulonephritis and nephrotic syndrome,and matched them with a same number of healthy control group by gender and age.An integrated learner based on the AdaBoost algorithm was used to establish a diagnostic model and assess its algorithm performance.JAVA was used to develop data presentation software.The accuracy of the AI model for disease judgment was assessed by manual verification using 199 abnormal urine tests.Results:Each report could be graded as four levels:normal,abnormal,ill and critical.Each item could be judged as normal,mild,moderate,severe or extreme and the population distribution was provided with big data.The training accuracy,true positive rate and area under the curve were≥88.3%,≥80.0%,and≥0.954 respectively using the machine learning model based on AdaBoost.The developed JAVA software presented the above results and displayed medical records and results,historical results,personalized advice,patient education and position in large population data.By manual verification,the accuracy rate of the AI model for disease judgment was 82.41%(166/199).Conclusion:This study established an intelligent interpretive reporting system for urine test results.It can distinguish the abnormality of each report,predict the disease of patients,and make personalized clinical decisions.