AHRNN:An Attention-based Hierarchical Recurrent Neural Network For Phenotype Classification

Currently,the broad adoption of Electronic Health Record(EHR)makes it pos-sible to access unprecedented amount of clinical data.Some relevant applications have improved the quality of health care and lowered the engendered cost.Phenotype classification,also known as “phenotyping”,is a relatively new yet critical medical informatics problem.The main goal is to classify patients by analyzing a series of EHRs including heart rate,blood pressure,etc.As precision medicine becomes an emerging approach to disease prevention and treatment,patients’ phenotypes also play an important role in triggering clinical decision support systems and predicting future health resource utilization.At the meantime,the research community is pursuing more and more advanced deep learning approaches,which have been proven effective and outperformed previous statistical or machine learning methods in extensive experiments from various domains.The development of deep learning approaches have brought important changes to human life.For instance,current machine translators,which employ end-to-end encoder-decoder recurrent neural networks,are more powerful than traditional rule-based machine translation paradigm.Players with actions controlled by Reinforcement Learning algorithms win not only computer players who depend heavily on brute force search,but also human masters in each game domain.Neural factorization machines as recommender systems provide more personalized results for users than conventional collaborative filtering methods.Models based on convolutional neural networks,such as DenseNet and ResNet,are capable to recognize images and detect objects more accurately than previous approaches with human crafted features.To let the technical improvement further benefit human welfare,this paper focuses on labeling phenotypes of patients in Intensive Care Unit(ICU)given their records from admission to discharge.Recent works mostly rely on Recurrent Neural Network(RNN)architectures to capture the temporal information for classification.Based on the assumption that the recurrent units are capable of remembering all the previous states,a prevalent practice for sequence representation is to leverage the last hidden state in the network.However,the hypothesis is too strong and falls short especially for tasks dealing with long sequences.Moreover,the memorizing strategy inside the recurrent neural networks is not necessarily capable of identifying the key health records for predicting phenotype labels.In this paper,we propose a novel hierarchical recurrent neural network with label-based attention mechanism for phenotype classification.Our intuition is to remember all the past records prior to prediction by a hierarchical structure and entitle the attention mechanism to extract crucial information from memory in the label’s perspective.To the best of our knowledge,it is the first work of applying attention-based hierarchical neural networks to clinical time series prediction.Experimental results demonstrate that our model is able to achieve higher accuracy with much lower time complexity than the state-of-the-art methods.