一种用于病案相似性度量的弱监督学习算法  

A Weakly Supervised Machine Learning Algorithm Applied to Similarity Measure of Medical Records

作　　者：张振宇 朱培栋[1,2] 赵东升 ZHANG Zhen-yu;ZHU Pei-dong;ZHAO Dong-sheng(School of Computer,National University of Defense Technology,Changsha 410073,China;School of Electronic Information and Electrical Engineering,Changsha University,Changsha 410022,China;Network Information Center,Academy of Military Medical Sciences,Beijing 100039,China)

机构地区：[1]国防科技大学计算机学院,湖南长沙410073 [2]长沙学院电子信息与电气工程学院,湖南长沙410022 [3]军事医学科学院网络信息中心,北京100039

出　　处：《计算机技术与发展》2019年第9期1-6,共6页Computer Technology and Development

基　　金：国家自然科学基金(61572514);长沙市科技计划重点项目(K1705007)

摘　　要：弱监督机器学习算法解决标签模糊类的问题具有更好的优势,该类算法缓解了数据标签的精度要求。病案的相似性度量就是这类问题,其对医疗数据的应用有着极其重要的基础性作用。鉴于现有病案相似性度量算法通常只基于病理关系的理论规则模型提出,忽略了数据本身包含的信息,文中提出一种弱监督机器学习算法应用于病案相似性度量。该算法首先基于多指标概率分配的方法进行病案组的构建,避免陷入局部最优的情况;然后根据理论模型进行标签赋值,充分利用理论信息;最后通过输入、损失函数、学习模型的分析,从机器学习的角度进行病案的相似性度量。与经典病案相似性度量算法相比,该算法提高了病案相似性度量的准确性,解决了高成本标签的问题。The weakly supervised machine learning algorithm has a better advantage in solving the label fuzzy class problem,which alleviates the accuracy requirements of data labels. The similarity measure of medical records is such a problem,which plays a significant role in medical applications. Given that existing medical records similarity measurement algorithm is usually based on the theoretical rule model under pathological relationship,this method ignores the information of the data itself. We propose a weakly supervised machine learning algorithm applied to the similarity measure of medical records. To start with,a medical record group is constructed based on a multi-index probability allocation method to avoid local optimal problems. In addition,the label assignment is conducted according to the theoretical model,which makes full use of the theoretical information. At last,through the analysis of input,loss function and learning model,the similarity measure of medical records is carried out from the perspective of machine learning. Compared with classical medical records similarity measurement algorithm,the algorithm proposed improves the accuracy and solves the problem of high cost labels.

关 键 词：弱监督 机器学习 病案相似性 理论模型 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]