| Electronic Health Records(EHR)are electronic data collections composed of disease diagnosis,medications,injections,laboratory test values,condition records and other medical information that occurred in patient’s hospitalization.In recent years,many studies have shown that the mining and secondary use of EHR data can promote the development of healthcare undertaking.However,due to the high dimensionality,sparsity,irregularity and complexity of EHR data,researchers face many challenges in the process of processing,exploring and applying EHR.Therefore,effective feature extraction and data representation are key steps before any further applications.In order to solve the above challenges,this thesis proposes two different patient representation learning methods based on Recurrent Neural Networks and attention mechanisms.The main research contents of this thesis are as follows:(1)Aiming at the unique hierarchical structure and the characteristics of EHR data,a multilayer representation learning method based on Bidirectional Long Short-Term Memory(BiLSTM)and attention mechanism is proposed.Firstly,the multi-head attention mechanism is used to analyze the potential correlation information in medical codes,and a linear transformation is added for non-negative vector learning,which aims to map the medical codes into non-negative realvalued representations.By summarizing the learned representations of the medical codes existed in visits,the initial visit vectors which are used as the inputs for the next part are obtained.Secondly,combining BiLSTM with self-attention mechanism to learn the weights of different visit vectors,and the patient representations are constructed by weighted summation.Finally,the learned patient representations are applied to patient’s mortality prediction task to evaluate the effectiveness of the proposed method.Experimental results demonstrate that the proposed method significantly improves the prediction effect of patient’s mortality.Furthermore,compared with raw data and other representation learning methods,the vector representations learned by the proposed method show their outstanding performance on multiple different classifiers.(2)Aiming at the contribution rates of EHR research tasks to data representation learning and the correlations between the tasks,a multi-task representation learning method based on Bidirectional Gated Recurrent Unit(BiGRU)and self-attention mechanism is proposed.First,BiGRU is used as an encoder to learn the hidden state vectors of patient’s visit sequences,which are used as the inputs for two different task networks of patient’s mortality prediction and sequence reconstruction.Secondly,these two different learning tasks are performed simultaneously to optimize the vector representations and improve the prediction performance of patient’s mortality.In the task of patient’s mortality prediction,a self-attention mechanism is introduced to measure the correlations between patient visit vectors and mortality,and different weights are learned for different visit vectors.At the same time,a fully connected layer with softmax classifier is incorporated for mortality prediction.Sequence reconstruction task is an unsupervised learning process,which uses BiGRU as a decoder to reconstruct the visit vectors.These two tasks share the encoding network and jointly contribute to the parameter optimization and feature learning of the network.Finally,the proposed method is applied to real EHR data set,and the experimental results show that the method is capable of learning effective patient representations and improving the predictive performance of patient’s mortality. |
