Application Of Deep Learning In Electronic Health Records Data

The emergence of deep learning has brought unprecedented development to such fields as machine vision and machine translation.In recent years,the combination of deep learning and medical data has attracted extensive attention,such as the application of deep learning in the diagnosis of diseases,prescription recommendation and the synthesis of Electronic Health Records(EHR).Therefore,this paper first reviews the existing deep learning research in EHR data and summarizes its achievements and shortcomings.Then studies were carried out on the generation of EHR data and the recommendation of prescriptions:First of all,there is patient privacy information in the EHR data,which severely restricts data sharing,resulting in lack of sufficient research data in related research;using the generated EHR data by Generative Adversarial Networks(GANs)can not only avoid data privacy problems well.Can also make up for the lack of data in actual research;Secondly,Based on the disease and corresponding prescription information in the EHR data,the Recurrent Neural Network(RNN)is used to construct a model reflecting the complex relationship between the disease and the prescription.For the complex disease combination,the corresponding prescription is recommended to assist the doctor in decision-making.This will greatly alleviate the lack of experience of grassroots doctors.Firstly,Grouped Correlational Generative Adversarial Networks(GcGAN)is proposed for the generation of EHR data.At present,existing algorithms for generating EHR data using GANs,such as ehrGAN and medGAN,regard the diseases,drugs,and treatments in EHR data as independent variables which are indiscrimin-ately entered into the model,without taking into account the meaning and grouping of them.Besides,the efficacy of treatment is often neglected.In view of the above problems,this paper first embeds the efficacy information into the disease representation,and then improves the Correlational Neural Network(CorrNet)to explicitly learn the intrinsic correlation between different groups of variables.Experiments on real EHR data demonstrate that the generated EHR data by GcGAN can achieve comparable performance to real data in terms of distribution statistics.At the same time,in the multi-label classification evaluation task,the introduction of generated data can boost the performance of multi-label classification and is superior to several most advanced methods.Finally,It can also automatically distinguishes between diseases,disease-specific drugs,and adjuvant drugs,which enhance the interpretability of the model to some extent.Secondly,MedAR is proposed for prescription recommendation.Prescription recommendation is a very important part after diagnosis.It can be regarded as a multi-label learning problem,in which RNN is used to achieve better results than traditional multi-label learning algorithms MLKNN and RAkEL.In this paper,according to the characteristics of disease diagnosis in EHR data,Word2vec was used to carry out distributed representation learning of diseases,and then attention was introduced to mimic the behavior of doctors to obtain an efficient visitor-level disease representation.Finally,we used the RethinkNet idea to enhance the correlation between the labels.Experimental results on real EJHR data show that MedAR can significantly improve the performance of prescription recommendations compared to methods such as DNN,Med_CNN,and CNN-RNN.