基于典型相关自编码器的过敏性鼻炎用药推荐

Medicine Recommendation for Allergic Rhinitis Based on Canonically Correlated Autoencoder

过敏性鼻炎患者的电子病历中存在着大量文本类型的主诉症状信息,其中蕴含了医生为患者做出诊断和医嘱用药的关键信息,而现有的药品推荐算法大多数只限于对患者的数值型、结构化数据的使用。针对这一问题,提出了一种基于深度典型相关自编码器的过敏性鼻炎用药推荐算法。首先通过一种基于搜索引擎的主诉文本结构化表示方法,从主诉文本中抽取症状标准信息;然后考虑患者的症状和用药之间存在较强的相关关系,通过构建深度典型相关自编码器模型对数据进行特征提取并且建立起主诉症状和用药情况之间的关联关系;最后根据患者的症状表征和用药表征通过加权近邻搜索进行药品的Top-N推荐。在一个真实的来自三甲医院耳鼻喉科的电子病历数据集上进行实验,验证了算法的准确性和有效性。

In the electronic health records of allergic rhinitis patients,there are a large number of text-type chief complaint information,which contains key information of doctor making diagnosis and prescribing medication for patients.However,most of the existing medicine recommendation algorithms only focus on the use of numerical and structured data of patients.To solve this problem,this paper proposes a medicine recommendation algorithm for allergic rhinitis based on deep canonically correlated autoencoder.Firstly,this paper extracts symptom standard information from the chief complaint through a structured representation method based on search engine.After that,considering the strong correlation between patients’symptoms and medication,a deep canonically correlated autoencoder model is built to extract the features of the data and establish the correlation between chief complaint symptoms and medication.At last,Top-N recommendation for allergic rhinitis is made according to the symptom representation and medication representation of patients.Experiments on a real electronic medical record dataset from the otolaryngology department of a first-class hospital demonstrate the accuracy and effectiveness of the algorithm.