基于多路径特征选择的发热待查分层分类辅助诊断方法  

An Auxiliary Diagnosis Method for Hierarchical Classification of FUO Based on Multi-Path and Feature Selection

作　　者：杜建超[1,2] 王燕宁 石磊 陈天艳[3] 梁婧晨[3] 王鑫 连建奇[4] 周云 Du Jianchao;Wang Yanning;Shi Lei;Chen Tianyan;Liang Jingchen;Wang Xin;Lian Jianqi;Zhou Yun(School of Telecommunications Engineering,Xidian University,Xi’an 710071,China;Guangzhou Institute of Technology,Xidian University,Guangzhou 510555,China;The First Affiliated Hospital of Xi’an Jiaotong University,Xi’an 710061,China;The Second Affiliated Hospital of Air Force Medical University,Xi’an 710038,China)

机构地区：[1]西安电子科技大学通信工程学院,西安710071 [2]西安电子科技大学广州研究院,广州510555 [3]西安交通大学第一附属医院,西安710061 [4]空军军医大学第二附属医院,西安710038

出　　处：《中国生物医学工程学报》2024年第6期682-692,共11页Chinese Journal of Biomedical Engineering

基　　金：空军军医大学第二附属医院前沿交叉研究项目(2021QYJC-005)。

摘　　要：发热待查病因种类多且特征维度高,令其难以准确诊断。本研究提出一种基于多路径特征选择的分层分类辅助诊断方法。首先,该方法根据发热待查病因的层次分类结构,设计了自顶向下的分层分类模型,在每一中间层择优选择可控数量的候选类别,构建多路径预测模式,并在多条路径中选择最终的最优分类;其次,在分层分类过程中采用L1,2范数正则化约束去除冗余特征,保留最优特征子集,提高预测准确性。此外,本研究收集西安交通大学第一附属医院自2011—2020年收治的发热待查患者就诊数据构建数据集,其中包含564条样本,327维特征,5个粗粒度类别:细菌感染、病毒感染、其他感染性疾病、自身免疫性疾病和其他非感染类疾病,以及下属的16个细粒度类别。在数据集上的十六分类验证结果表明,当所提方法在中间路径数目为3、选择25%的特征时,综合预测准确率达到76.08%,FH指标达到86.72%,FLCA指标达到85.39%,比传统的单路径、不进行特征选择的方法分别高出9.42%、4.69%和3.36%。相比扁平化分类算法和现有其它同类分层分类算法,所提方法的性能指标均有明显提升。本模型的提出为发热待查辅助诊断提供了新方法。Many causes of fever of unknown origin(FUO)and high characteristic dimensions lead to difficulty in accurate diagnosis.This paper proposed an auxiliary diagnostic method based on hierarchical classification with multi-path and feature selection.Firstly,according to the structure of FUO causes,this method designed a top-down hierarchical classification model to select a controllable number of candidate categories in each middle layer,constructing a multi-path prediction mode,and finally selecting the optimal classification among multiple paths;secondly,an L1,2 paradigm regularization constraint was utilized to eliminate redundant features and preserve the optimal subset of features to reduce interference and improve prediction accuracy.In addition,this paper collected data from the First Affiliated Hospital of Xi'an Jiaotong University regarding patients visiting for FUO from 2011 to 2020 to construct a comprehensive dataset.This dataset included 564 samples and 327 dimensional features,categorized into five coarse-grained categories:bacterial infections,viral infections,other infectious diseases,autoimmune diseases,and other non-infectious diseases,and into 16 subordinate fine-grained categories.The sixteen-classification verification results on the dataset showed that when the proposed method selected 25%of the features with 3 candidate classes in the middle layer,the accuracy,FH and FLCA reached 76.08%,86.72%and 85.39%,respectively,which were 9.42%,4.69%,and 3.36%higher than the traditional single-path and non-feature selection methods,respectively.The proposed method significantly improved evaluation performance compared to the flat classification algorithms and other existing hierarchical classification algorithms,providing a more effective auxiliary diagnostic method for FUO.

关 键 词：发热待查 智能诊断 机器学习 分层分类 特征选择 

分 类 号：R318[医药卫生—生物医学工程]