Dipole: Diagnosis Prediction in Healthcare via Attention-based Bidirectional Recurrent Neural Networks

摘要

Predicting the future health information of patients from the historicalElectronic Health Records (EHR) is a core research task in the development ofpersonalized healthcare. Patient EHR data consist of sequences of visits overtime, where each visit contains multiple medical codes, including diagnosis,medication, and procedure codes. The most important challenges for this taskare to model the temporality and high dimensionality of sequential EHR data andto interpret the prediction results. Existing work solves this problem byemploying recurrent neural networks (RNNs) to model EHR data and utilizingsimple attention mechanism to interpret the results. However, RNN-basedapproaches suffer from the problem that the performance of RNNs drops when thelength of sequences is large, and the relationships between subsequent visitsare ignored by current RNN-based approaches. To address these issues, wepropose {\sf Dipole}, an end-to-end, simple and robust model for predictingpatients' future health information. Dipole employs bidirectional recurrentneural networks to remember all the information of both the past visits and thefuture visits, and it introduces three attention mechanisms to measure therelationships of different visits for the prediction. With the attentionmechanisms, Dipole can interpret the prediction results effectively. Dipolealso allows us to interpret the learned medical code representations which areconfirmed positively by medical experts. Experimental results on two real worldEHR datasets show that the proposed Dipole can significantly improve theprediction accuracy compared with the state-of-the-art diagnosis predictionapproaches and provide clinically meaningful interpretation.