Research And Application Of Improved LSTM For Electronic Health Record

Electronic health records(EHR)are a collection of data such as patients' medical history,diagnosis,medications,treatment plans and radiographic images.Therefore,it is valuable to improve the patients' healthcare plan by mining EHRs.However,due to the irregularity and sparseness of EHRs,research of EHRs faces many challenges.In recent years,with the successful application of deep learning technology in the field of medical information,the effect of analyzing EHRs has been greatly improved.This paper summarizes and analyzes the existing researches of EHRs.In view of the shortcomings of the existing researches,which can't effectively mine the temporal characteristics between multiple visits and the key factors that affect the patients' disease.This paper proposes two prediction methods based on Long Short-Term Memory(LSTM)which realizes the prediction of mortality and disease category of patients during hospitalization based on EHRs.The main work and innovations of the paper are summarized as follows:(1)A method for predicting patients' mortality during hospitalization based on an improved LSTM with training strategy by target replication is proposed.Firstly,medical diagnosis codes(ICD-9)during patients' hospitalization are extracted from the MIMIC-III database,and then a diagnostic codes weight method is proposed to calculate the impact of the historical diagnosis.The weighted feature matrix is normalized further to serve as the input of an improved Long Short-Term Memory(LSTM*).Secondly,LSTM* is constructed via removing input and output gates.Also,an effective training strategy is proposed to train LSTM* to predict patients' mortality during their hospitalization.At last,experiments results demonstrate that the proposed model not only effectively increases the prediction accuracy of patients' mortality but also reduces the calculation time compared with other methods.(2)A disease category prediction method based on BiLSTM combined with attention mechanism is proposed.Firstly,the laboratory data,physiological indicators and diagnosis time during the patients' hospitalization are extracted from the MIMIC-III database.Then,the patients' representation vector was learned through the BiLSTM combined with attention mechanism,and finally predicted the disease category of the patient suffered during hospitalization.Experiments show that compared with other methods,the proposed prediction method effectively improves the accuracy of disease category prediction.