Research On Prediction Methods Based On Recurrent Neural Network And Their Applications

With the continuous development of the medical industry and information technology,the electronic health record(EHR)system has been widely used,accumulating a large amount of EHR data,which contains a lot of valuable information.Researchers can build prediction methods through analysis of EHR data to make the medical services of the EHR system more intelligent and efficient,but the prediction of clinical results on the EHR dataset faces many challenges,such as mixed medical data formats,high dimensionality,and related fields.Previous research methods mainly used traditional machine learning methods,which relied heavily on manual feature extraction,which required a lot of time and effort.In addition,deep learning models lacked interpretability.To solve these challenges,two prediction methods are proposed based on deep neural networks.The main work of this paper is as follows:(1)In order to solve the high-dimensional and time-dependent problems of EHR data,the paper proposes a prediction method for hospital mortality based on Attention-LSTM.Firstly,use the skip-gram method in word2 vec to vectorize the physiological indicators in EHR,and then input the generated vectors into a trained improved LSTM network,and combine the attention mechanism to predict the patient’s mortality.Testing on the MIMIC-III dataset,the experimental results show that the proposed method has obtained more accurate prediction results than other methods,and the weights generated during the experiment are analyzed,which proves the importance of adding attention mechanism.(2)In order to solve the problem of future contextual information ignored by the RNN method,the paper proposes a clinical event prediction method based on bidirectional LSTM.Firstly,the skip-gram method in doc2 vec is used to vectorize the disease code in the EHR.The generated vector is input into the bidirectional LSTM,and the weight of each diagnosis code is calculated in conjunction with the attention mechanism,and then input to the bidirectional LSTM.Calculate the weight of each diagnosis with the attention mechanism,and finally predict the mortality of patients and the sequence of diagnosed diseases.The advantage of the bidirectional LSTM is that it can provide complete past and future context information for each diagnosis in the input sequence of the output layer.Just like a doctor looking at a patient’s previous diagnosis record when diagnosing a patient,it can better diagnose the patient’s condition.Finally,the test is performed on the MIMIC-III data set.Experimental results show that the proposed method can obtain more accurate prediction results,and the interpretability of the model is proved in visualization.