| With the rapid development of scientific technology and medical services,the diagnosis and treatment of diseases have become increasingly scientific and standardized.However,medical diagnosis at this stage mainly depends on the doctor’s professional knowledge and clinical experience.With the development of medical health business,the scale of health care data is also getting larger and larger.From these data,information on diseases can be tapped.The electronic health record itself contains a wealth of patient-related information,such as demographic information,illness and medication status.The information will help predict the patient’s future illness,and then provide diagnostic aids for medical decisions."The Outline of the "Healthy China 2030 Plan"" also clearly points out that to promote the construction of healthy China,we must adhere to prevention,and strengthen early diagnosis,early treatment,and early recovery.The study of diagnosis prediction can solve the above problems and improve medical services.At the same time,diagnosis prediction is key technology for precision medicine.But there are still some problems that have not been well resolved.The electronic health record itself has a large amount,a high dimension,the presence of missing and noise,and timing characteristics.These characteristics make it difficult to use traditional machine learning algorithms to solve diagnosis prediction problems.Some people have applied for deep learning models such as recurrent neural networks to disease diagnosis prediction and have achieved good results.However,there are still some problems.For example,electronic health record data is not fully utilized and performance of the model is not ideal.According to the current research,the specific work and contributions of this paper are summarized as follows:1.A diagnosis prediction model "RNN-INFO" based on demographic information and recurrent neural networks is proposed.The model uses the diagnosis and medication status within the observation window to effectively predict the patient’s probability of illness and disease diagnosis in the prediction window.The model first uses a fully connected layer to reduce the dimension of the input data,and then uses a two-layer recurrent neural network to capture the hidden representations of the electronic health record.In order to further improve the accuracy of the model,the model also introduces the patient’s demographic information.Through the experimental results,the model is better than other baselines and RNN models that do not include demographic information.2.A diagnosis prediction model "RNN-ATTEN" based on attention mechanism and recurrent neural networks is proposed.The model uses the patient’s historical medical sequence to predict next visit.The attention mechanism is added to better attention to the visit that has a significant impact on the diagnosis of the disease,so that the model has a better effect.The input of the model is the patient’s medical information sequence,and the input data is obtained through the two recurrent neural networks to obtain the patient’s medical hidden identity and the patient’s medical identity to represent the attention.Then combine the two to predict the patient’s last visit diagnosis.Experiments show that experimental results based on the attention mechanism model are indeed better than other models.Finally,we study the interpretability of the model.The two models we propose conduct experiments and verifications on medical insurance medical dataset in a certain place of northern China.For the "RNN-INFO"model,after adding demographic information,the model significantly improves the probability of medical care and the diagnosis of the medical visit.For the "RNN-ATTEN"model,the performance of the model is better than the model that does not introduce the patient’s attention mechanism.Therefore,two deep learning models proposed in this paper are helpful for the improvement of disease diagnosis prediction. |
PDF Full Text Request