Abstract—
The prediction of patient’s future health
information from the historical electronic health records (EHR)
forms the core of the development of personalized healthcare
research tasks. Patient EHR data consists of sequences of visits
over time, where each visit contains multiple medical codes,
including diagnosis, medication, and patient profile. Using
historical data from the EHR, we can predict medical
conditions and medication uses. Existing works model EHR
data by using recurrent neural networks (RNNs). However,
RNN-based approaches have certain limitations: the
performance of RNNs drops when the length of sequences is
large and they ignore some of the characteristics of the patients
themselves. We propose an application of using bidirectional
RNNs to remember all the information of both the past and
future visits and add some patient’s characteristics as side
information into this model. Experimental results on real
world EHR datasets show that the proposed model can
remarkably improve the prediction accuracy when compared
with the diagnosis prediction approaches, and it can provide
clinically meaningful interpretation.
Index Terms—Component, electronic health records,
bidirectional recurrent neural networks, side information
I. INTRODUCTION
The common challenge in smart health is how to use the
large amount of data in predicting visiting patients’ diseases
in a short period of time. Due to complicated processes,
different symptoms, and pathological tests, making the
correct diagnosis is a difficult task and causes delays in
providing the proper treatment. Electronic health records
(EHR) consisting of patient health data, including
demographics, diagnoses, procedures, and medications, have
been utilized successfully in several predictive modeling
tasks in healthcare [1]-[3]. EHR data are temporally
sequenced by patient medical visits that are represented by a
set of high dimensional clinical variables (i.e., medical
codes). While forecasting medical models have been
developed to predict the expected demand, most of the
existing works have focused on specialized forecasting
models or a single target. In order to model the sequential
EHR data, recurrent neural networks (RNNs) are used in the
literature to obtain accurate and robust representations of
patient visits in diagnostic predictive tasks [3], [5].
However, the predictive power of these models drops
significantly when the length of the patient visit sequences is
large. Further, these models usually ignore some of the
Manuscript received January 8, 2018; revised March 19, 2018.
The authors are with College of information Science &Technology,
Hainan University, Haikou, China and State Key Laboratory of marine
resource utilization in the South China Sea, Hainan University (e-mail:
muyangzi521@163.com, huangmx09.com, cyye@ustc.com,
wuqingzhou@21cn.com).
characteristics of the patients themselves and others. While
not so extreme, there are many diseases associated with
gender, family history, region, season, and so on.
Bidirectional recurrent neural networks (BRNNs) [6], which
can be trained using all the available input information in
the past and future, have been used to alleviate the problem
of long sequences, thereby improving the predictive
performance. Referring to the method of collaborative
filtering (CF), we use the side information to reasonably
interpret the importance of patients and medical codes in the
prediction results. This side information can be obtained
from the user profile and other information. Some side
information has proven to be useful for heart disease
decisions [7], [8]. Some hybrid CF methods have gained
popularity in recent years [9], [10], where side information
is integrated into matrix factorization to learn the effective
latent factors.
We demonstrate that the proposed model achieves
significantly higher prediction accuracy when compared to
the other approaches in diagnosis prediction, using our
datasets from Haikou People’s Hospital. In summary, our
main contributions are as follows:
We propose a new, end-to-end, simple, and powerful
model that can accurately predict future visits, without
relying on any expert’s medical knowledge.
It models the patient’s visit information in time- and
reverse-time-ordered ways and employs side information
as supplementary information.
We show that the proposed new model outperforms
existing methods in diagnosis prediction with regard to
EHR datasets.
The rest of this paper is organized as follows: In Section
II, we discuss the connection between the proposed
approaches and related works. Section III details the
proposed new model. The experimental results are given in
Section IV. Section V concludes this paper.
II. RELATED WORKS
This part reviews the existing work for mining EHR data.
In particular, it focuses on several state-of-the-art models on
diagnosis prediction tasks. It also includes some works that
use side information in CF-based methods.
A. EHR Data Mining
Mining EHR data is a popular topic in medical informatics.
The investigated tasks include electronic genotyping and
phenotyping [11], [12], disease progression [13], [14],
diagnosis prediction [1], [2], [15], and so on. In most of these
tasks, the machine learning model and depth neural network
models can significantly improve the performance.
Diagnosis prediction is an important and difficult task in
medical informatics. Machine learning can remarkably
improve performance, such as using the SVM algorithm in
Diagnosis Prediction via Recurrent Neural Networks
Yangzi Mu, Mengxing Huang, Chunyang Ye, and Qingzhou Wu
International Journal of Machine Learning and Computing, Vol. 8, No. 2, April 2018
117
doi: 10.18178/ijmlc.2018.8.2.673