Interpretable Predicting Methods Of Clinical Events Based On Knowledge Graph

Electronic Health Records(EHR)are collections of a series of clinical events generated by patients during one or more visits.With the development of artificial intelligence and the popularization of computer technology,the traditional paper health records have been gradually replaced with the electronic health records in the medical field.Researchers are conducting research based on the existing electronic health record data,but they also have shortcomings in the research.On the one hand,the electronic health records data is a time series data,including the patient’s medical treatment time,which has the problem of irregular time intervals.However,some rich time information has been ignored by many known studies;On the other hand,the interpretability of deep learning models is poor,although deep learning has been widely used in the predictions of the clinical events and performed well.In the medical field,the interpretability is always very important to users.In response to above problems,the paper launches research on predicting methods for interpretable clinical events.Which includes modeling for electronic health record data to capture the patient’s time interval information.Using model-independent methods to analyze the prediction results.The main work and innovations of this paper are as follows:(1)Propose a prediction method based on attention mechanism and time-aware long short term memory network.This method can capture the time interval information during the patient’s consultation process and focus on the key information which is useful for the current prediction.Firstly,by defining the time attenuation function,the time interval in the patient’s visit sequence is converted into weights.The converted weights are used to calculate the information that the previous memory unit should be retained to achieve the improvement of the previous memory unit;Secondly,the attention mechanism is introduced to process the input data,which aim to assign different features with different weights to improve the prediction ability of the model.Finally,a large number of experiments are carried out on the MIMIC data set.The experimental results show that the model can capture the time information in the patient’s visit sequence and the event information.What’s more the model’s predictive performance is better than other comparison methods.(2)Propose an interpretable prediction method based on ensemble learning.This method combines the advantages of ensemble learning and deep learning,which can improve the prediction performance of the model and provide the interpretability of the results.Firstly,use the idea of ensemble learning to continuously train weak classifiers,a strong classifier will be obtained from the weak classifiers which have been made a weight fusion.According to the prediction result of the strong classifier,the input sequence will be reconstructed and input it into the deep learning model to obtain the final prediction results;Secondly,use the method of feature importance to analyze the interpretability of the prediction results of the model,and explore the impact of different features on the prediction results;Finally,a large number of experiments on the MIMIC data set show that the model improves the accuracy of the prediction.The interpretability of the results can play a good auxiliary role in predicting clinical events.(3)Propose an interpretable prediction method for the knowledge graph of pneumonia.This method can dynamically mine the influence of different feature sets on the prediction results,and can improve the prediction performance of the model for the interpretability of the results.Firstly,construct a pneumonia knowledge graph based on the information in the MIMIC data set,and make a model based on the constructed knowledge graph;Secondly,vectorize the triples in the pneumonia knowledge graph,and combine reinforcement learning with gated recurrent units to predict the patient’s future Prevalence;Finally,combining the results of reinforcement learning and expert knowledge to analyze the impact of different feature sets on the prediction results of the model.Then according to the analysis results,the corresponding features are selected and recombined to make predictions.The experiment shows that the proposed method improves the predictive performance of the model and it has a strong practical value.