引入注意力机制的在线问诊推荐研究  

Online Doctor Recommendation System with Attention Mechanism

作　　者：聂卉[1] 蔡瑞昇 Nie Hui;Cai Ruisheng(School of Information Management,Sun Yat-Sen University,Guangzhou 510006,China;The Seventh Affiliated Hospital,Sun Yat-Sen University,Shenzhen 518107,China)

机构地区：[1]中山大学信息管理学院,广州510006 [2]中山大学附属第七医院,深圳518107

出　　处：《数据分析与知识发现》2023年第8期138-148,共11页Data Analysis and Knowledge Discovery

基　　金：2020广州市科技计划项目(项目编号:202002020036)的研究成果之一。

摘　　要：【目的】将深度学习引入医疗推荐领域,解决在线诊疗过程中患者“择医”面临的实际问题。【方法】利用患者问诊记录,采用层次注意力网络(HAN)构建医生与患者模型;设计基于“医患”适配度和患者“评分”的医生推荐方案。两个方案应用深度学习框架HAN构建医生和患者模型,并运用注意力机制加强“医患”向量间的交互,使医生名下的与求诊者病情相似的患者获得更高权重,据此计算医生推荐值。【结果】HAN能够从患者的疾病描述中提取表征病情的关键信息,通过提升建模质量,推荐命中率相较经典的Word2Vec模型提升了16.45个百分点;对于推荐值计算,基于注意力机制的“评分”方案的命中率最高(79.7%),显著优于基于余弦相似度的计算方案(74.9%)。【局限】仅利用医生名下历史患者的问诊数据为医生建模,医生的口碑、资历、专长等信息未纳入模型。【结论】构建用户和推荐对象模型是设计推荐系统的关键,增强用户和推荐对象之间的特征交互可以提高推荐质量,本研究验证了基于深度学习的建模技术在推荐任务中的优势。[Objective]This paper utilizes deep learning to recommend medical services for patients,which helps them choose doctors during online diagnosis and treatment.[Methods]First,we used the Hierarchical Attention Network and patient consultation records to construct doctor-patient models.Then,we designed doctor recommendation schemes based on the“doctor-patient”compatibility and patient“rating”.Both schemes incorporated the HAN deep learning framework to build doctor-patient models and used attention mechanisms to enhance the interaction of“doctor-patient”.Patients with similar conditions to those inquiring about treatments receive higher weights,which helped us calculate the doctor’s recommendation score.[Results]The HAN model could extract the critical information representing the patient’s condition from their disease descriptions.The recommendation hit rate was improved by 16.45%compared to the classical Word2Vec model by improving the modeling quality.For the recommendation score,the“rating”scheme based on the attention mechanism achieved the highest hit rate(79.7%),which is significantly outperforming the cosine similarity-based scheme(74.9%).[Limitations]This study only utilized historical patient consultation data under each doctor’s name to model the doctors,and the model did not include information such as the doctor’s reputation,credentials,and expertise.[Conclusions]Constructing user and recommendation objects is crucial in designing recommendation systems.Enhancing feature interaction between the users and recommendation objectives can improve recommendation quality.This study validates the advantages of deep learning modeling techniques in recommendation tasks.

关 键 词：推荐系统 智慧医疗 深度学习 在线问诊 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]