See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/330876062

Deep Learning Techniques for Electronic Health Record Analysis

Conference Paper · July 2018

DOI: 10.1109/IISA.2018.8633647

CITATIONS READS
2 94

4 authors:

Luciano Caroprese Pierangelo Veltri

Università della Calabria Universita' degli Studi "Magna Græcia" di Catanzaro
42 PUBLICATIONS   228 CITATIONS    249 PUBLICATIONS   1,500 CITATIONS   

SEE PROFILE SEE PROFILE

Eugenio Vocaturo Ester Zumpano

Università della Calabria Università della Calabria
14 PUBLICATIONS   32 CITATIONS    99 PUBLICATIONS   828 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Computational Techniques for Biological Data Analysis View project

Electroacoustic voice analysis View project

All content following this page was uploaded by Eugenio Vocaturo on 28 June 2019.

The user has requested enhancement of the downloaded file.

 Deep Learning Techniques
for Electronic Health Record Analysis
Luciano Caroprese1 , Pierangelo Veltri2 , Eugenio Vocaturo1 and Ester Zumpano1
1 Università della Calabria

87030 Rende, Italy

e mail: [l.caroprese|e.vocaturo|e.zumpano]@dimes.unical.it
2 Università Magna Graecia di Catanzaro

88100 Catanzaro, Italy

E mail: [email protected]

Abstract—Following the general trend, the amount of digital reporting these is out of the scope of this paper. Bottom line,
information in the stored electronic health records (EHRs) had an the following seven V’s provide a good definition of Big Data.
explosion in the last decade. EHRs are not anymore used, as in the
past, to store basic information of the patient and administrative • Volume. Volume is the total amount of data. In the recent
tasks, but they may include a range of data, including the medical past it was measured in Gigabytes, now Zettabytes (ZB)
history of the patient, laboratory test results, demographics, or even Yottabytes (YB) are used.
medication and allergies, immunization status, radiology images,
vital signs. At the present, the problem has shifted from collecting
• Velocity. Velocity is the speed in which data is produced,
massive amounts of data to understanding it, i.e. use EHRs processed and analyzed. Making quick decision is crucial
for turning data into knowledge, conclusions and actions. EHRs for a company and avoids stasis.
were not designed to forecast disease risk or disease progression • Variety. Variety describes the different format of data that
or to determine the right treatment, but if they are combined cannot be stored in structured relational database systems.
with artificial intelligence (AI) algorithm this issue became
possible. The need for tools allowing to construct predictive
Organizing rapidly changing data having different format
models capturing disease progression is a priority. In the recent is one of the biggest challenge of Big Data.
past EHRs were analyzed using traditional machine learning • Variability. Variability is different from variety. It has
techniques, whereas recently the progress in the field of deep several potential meaning. Is the data consistent in terms
learning let to the application of deep learning techniques to of availability at times? In the presence of extreme values
EHRs. This paper reports a brief overview of some recently
developed deep learning tools for EHRs.
are these outliers or just noisy data?
Index Terms—Electronic Heath Record, Deep Learning, Pre- • Veracity. Veracity refers to the accuracy of the data.
dictive Tools. Data have to conform with truth, therefore key questions
are related to the provenance of data, to the reliability,
I. I NTRODUCTION accuracy and completeness of sources.
recent study of the IBM [1] states that about the 90%
A of the data in the world has been created in the last two
years [1]. The current output of data is roughly 2.5 quintillion
• Visualization. Visualization refers to the need of tools
to figure out data and analysis’ results on data. Having a
vast amount of data is useless if you are not effective in
bytes a day. The IDC white paper by Seagate [2] forecasts that capturing the meaningful part of it.
data creation will swell to a total of 163 zettabytes (ZB) by • Value. Value is the final aim of Big Data. Each organi-
2025. Seagate’s report also predicts that in the next future the zation has to get value from the data.
primary creators of the world’s data will not be consumers, but
Following the general trend, the amount of digital informa-
enterprises that will create the 60% of the world’s data in 2025.
tion in the stored EHRs had an explosion in the last decade.
This indicates that the very next will be an era focused on the
EHRs are not anymore used, as in the past, to store basic
value of data: creating, utilizing and managing life critical data
information of the patient and administrative tasks, but they
necessary for the smooth running of daily life for consumers,
may include a range of data, including the medical history of
governments and businesses. Big Data are everywhere! They
the patient, laboratory test results, demographics, medication
are part of the information processing system of all sectors,
and allergies, immunization status, radiology images, vital
from science to government, from health care to media, from
signs.
university to real time commerce, from security/privacy to
financial [3], [4], [5]. But what is big data? Many different At the present, the problem has shifted from collecting
definitions have been provided in the recent literature and massive amounts of data to understanding it, i.e., turning data
into knowledge, conclusions, and actions. One of the major
task of modern medicine is the possibility of anticipating the
978-1-5386-8161-9/18/$31.00 2018
c IEEE future evolution of a disease, by predicting the risks for the
patient so that helping in prevention. This specific task is often Neural Network (FFN) in which the information moves in
referred as prospective health and is of great relevance in the only one direction, forward, from the input nodes, through
more general task of personalized medicine. The need for tools the hidden nodes (if there exist) and finally to the output
allowing to perform prediction is a priority. nodes. There are no cycles or loops in the network. Deep
In the recent past EHRs were analyzed using traditional learning algorithms are built upon Artificial Neural Network
machine learning techniques, whereas recently the progress in (ANN) in which a number of interconnected nodes (neurons)
the field of deep learning let to the application of deep learning are organized in a number of different layers. Hidden nodes
techniques to EHRs. are those not belonging to the input or output layers.
The interest from both industries and academia on the use
of deep learning on EHR at different level and with different
specific tasks is testified by the effort devoted by the general
research community on this relevant reseach topic, by the
number of proposed applications and by the general interest on
the topic by public media, premier conferences and prestigious
journals. The amount of literature in this domain is huge and
revise it is out of the scope of the paper.
This paper reports a brief overview of some recently devel-
oped deep learning tools for EHRs.
II. BACKGROUND
a) Electronic Health Records (EHR): The term Fig. 1. NN with 1 input layer, 2 hidden layer and 1 output layer
electronic health record (EHR), or electronic medical record
(EMR), refers to the collection of patient health information
III. T OOLS FOR EHR
in a digital format. They can be categorized in terms of
functionalities: (i) basic EHRs without clinical notes, (ii) basic A. Deepr
EHRs with clinical notes and (iii) comprehensive systems. The paper in [6] presents Deepr, a deep learning system that
EHRs, even in its simple form, provide a rich collection learns how to extract features medical records with the final
of data for the researchers. Data can be shared across the aim of predicting risk for the patient. Deepr is a multi layered
network and may include, as previously specified, a range of architecture based on convolutional neural nets (CNNs). A
information. EHRs were primary designed for internal hospital medical record is a sequence of visits and for each visit there
administrative tasks and many different schemata exist in is a subset of coded diagnosis, lab tests and textual text. A
different structure. Therefore, one of the major challenge young person is expected to have few medical records, whereas
consists in integrating and harmonizing data belonging to an old person is expected to have more medical records.
different institutions. In addition the heterogeneous nature of Medical records are episodic, i.e. they are recorded when
the data types including numerical data, date time objects, patient visit clinics and the time interval between two visit
free text etc poses significant challenges in working with is irregular. Therefore, medical records are irregular at patient
EHRs. level. Anyhow, data are regular at local episode level and these
regularities can be thought of as clinical motifs. Deepr stems
b) Deep Learning: A Neural Network (NN) can be from this observation: the key point is being able to identify
represented by a directed acyclic graph in which the input regular clinical motifs of irregular data.
layer takes in signal vectors and one or multiple hidden layers In order to describe the system we refer to Fig. 2, provided
process the outputs of the previous layer. Deep learning model in [6]. In the following steps are reported from left-to-right.
is the approach aiming at producing an end-to-end system • The top-left box in the figure represents a generic exam-
that learns from raw data and performs specific tasks without ple of a medical record consisting of multiple visits each
manual supervision. A Deep Neural Network presents many having a number of associated coded objects (procedures
more layers and many more nodes in each layer with respect and diagnosis). The risk of the patient is unknown and is
to a neural network. The number of parameters to be tuned represented in the box by a question mark.
increases with respect to those of a neural network and, in • The medical record is transformed into a “sentence” of
addition, they cannot be learned without enough data and multiple phrases (each phrase represents a visit), sepa-
without powerful computers. rated by “special words” that denote the time gap between
There exists different types of deep neural architectures such two visits. This allows to represent a medical record as
as: (i) the Convolutional Neural Network (CNN) in which a sentence containing information of all important events
the connectivity pattern between layers is inspired by the whose length is variable.
organization of the visual cortex; (ii) the Recurrent Neural • The other layers of Deepr constitutes a CNN. At the first,
Network (RNN) where connections between nodes form a words are embedded into a continuous vector space, next,
directed graph along a sequence and (iii) the Feedforward using convolution operators, local motifs are detected.
 Fig. 2. Deepr Architercture Overview

• Local motifs are then pooled to obtain a global feature

vector, that is passed to a classifier in order to predict an
output, that is the future risk.
Deepr has been validated over a database of 300k patients
of an Australian hospital. The focus of the experimentation
was predicting unplanned readmission within 6 months. It
demonstrated a superior accuracy with respect to existing bag-
of-words representation and the capability to learn predictive
clinical motifs.

B. Deep Patient Fig. 3. Deep Patient: Conceptual framework used to derive the deep
patient representation from EHR data warehouse
The work in [10] proposes the Deep Patient model: an
unsupervised deep learning approach to extract the hierarchical
features and patterns from EHRs data. The paper shows that
unsupervised deep learning can be profitable used for obtain-
ing the underlaying hierarchical features of EHRs datasets.
In order to describe the system we refer to Fig. 3 and Fig.
4 provided in [10]. Figure 3 reports the high level conceptual
framework used to obtain the Deep Patient representation.
EHRs are extracted from the hospital data warehouse, after
they are pre-processed to derive clinically relevant phenotypes
and are grouped in patient vectors (i.e., raw representation, Fig. 4. Deep Patient: Diagram of the unsupervised deep feature
Fig.3A). Each patient is represented by just a single vector or learning pipeline
by a sequence of vectors. The set of vectors, obtained from
all the patients, is used as input of the unsupervised deep
feature learning algorithm that derives high level descriptors model, the disease prediction is evaluated in two different
(Fig.3B). Each patient of the data warehouse is represented clinical tasks: disease classification (i.e., evaluation by disease)
using the same features, in addition this kind of deep represen- and patient disease tagging (i.e., evaluation by patient).
tation can be applied to different clinical tasks (Fig.1C). The The Deep Patient model has been experimented over the
unsupervised deep feature learning pipeline that transforms a Mount Sinai data warehouse which collects emergency room,
row dataset into a deep patient representation is reported in inpatient and outpatient visits.
Fig. 4. The diagram reports multiple layers of neural networks
in which each layer is trained in order to derive a higher- C. Med2Vec
level representation with respect to the output of the previous Med2Vec [7] is a simple and robust algorithm that is able
layer. The last network of the chain outputs the final patient to efficiently learn code and visit level representation by using
representation. This process, that is similar to neuroscience real world EHRs datasets, avoiding the need of expert medical
models of cognition, combines lower-level features into a knowledge. Med2Vec uses a two level neural network for
unified and more compact representation. learning lower dimension representations of clinical concepts
The Deep Patient framework essentially uses stacked de- from large EHRs datasets with over million visits and allows
noising autoencoders (SDA) to learn the data representations to interpret the learned representation.
from multi-domain clinical data. After the construction of the In order to describe the system we refer to the architecture
in Figure 5 provided in [7]. The following description reports
bottom-up steps.

Fig. 6. Doctor AI Overview

Doctor AI extracts clinical events from EHRs datasets and

learns a model. Given a new patient’s record, it is able to fore-
cast the patient’s diagnoses (Dx), the prescribed medications
(Rx) and it is able to predict when the patient will make the
next visit.
IV. C ONCLUDING R EMARKS

The analysis of EHRs is an ongoing research area, that

if correctly managed, can offer a great opportunity to
Fig. 5. Med2Vec Architercture Overview
construct predictive models capturing disease progression.
Many challenges wait for a solution and the interest of the
The set of medical codes in the EHRs dataset c1 , c2 , ...c|C| research community around this relevant and intriguing topic
is denoted by C and has size |C|. Given a visit Vt the is certified by the increasing number of publications related to
corresponding visit representation is obtained by using a multi- Deep EHRs. The wide spread of EHRs and the vast amount
layer perceptron (MLP). Firstly, the visit Vt is represented by of collected information urgently requires efficient tools for
a binary vector xt ∈ {0, 1} in which the i − th visit is 1 only turning data into knowledge, conclusions and actions. This
if ci ∈ Vt . Then xt is converted into an intermediate visit paper reports a brief overview of some recently developed
representation ut . The demographic information dt , contained deep learning tools for EHRs.
in EHRs, is concatenated with ut in order to create the final
Name Task
visit representation vt which is trained in order to predict its Deep Patient EHR representation. Multi-outcome prediction
neighboring visits ..., xt−2 , xt−1 , xt+1 , xt+2 ... Med2Vec EHR representation
The performance of Med2Vec have been evaluated on a Deepr Hospital re-admission prediction
dataset provided by Childrens Healthcare of Atlanta (CHOA) Doctor AI Multi-outcome prediction
and on the CMS dataset, a publicly available synthetic med-
ical claims dataset. Results show Med2Vec improves predic-
R EFERENCES
tion accuracy, but also identifies influential clinical concepts.
The source code of of Med2Vec is publicly available at [1] https://www-01.ibm.com/common/ssi/cgi-
bin/ssialias?htmlfid=WRL12345USEN
https://github.com/mp2893/med2vec. [2] D. Reinsel, J. Gantz, J. Rydning. Data Age 2025: The Evolution of
Data to Life-Critical Don’t Focus on Big Data; Focus on the Data
That’s Big IDC White Paper. https://www.seagate.com/www-content/our-
D. Doctor AI story/trends/files/Seagate-WP-DataAge2025-March-2017.pdf
[3] S. Sohangir, D. Wang, A. Pomeranets, T. M. Khoshgoftaar. Big Data:
Doctor AI [9] is an application of RNN used to forecast Deep Learning for financial sentiment analysis. J. Big Data 5: 3 (2018).
future disease diagnosis and medical prescriptions. The basic [4] G. Halevi, H. Moed. The evolution of big data as a research and scientific
idea of Doctor AI is to learn an effective representation of the topic: Overview of the literature, Res. Trends 3-6 (2012)
[5] G. Lansley, P. Longley. Deriving age and gender from forenames for
medical record of the patient over the time in order to predict consumer analytics. Journal of Retailing and Consumer Services, 30, 271-
future clinical events. 278 (2016).
Considering the fact that medical records have different [6] P. Nguyen, T. Tran, N. Wickramasinghe, S. Venkatesh. Deepr: A Con-
volutional Net for Medical Records. IEEE J. Biomedical and Health
length and that RNN have been shown to be successful in Informatics 21(1): 22-30 (2017).
representation learning of sequential data they have been [7] E. Choi, M. Taha Bahadori, E. Searles, C. Coffey, M. Thompson, J. Bost,
chosen for implementing Doctor AI. The authors state that, J. Tejedor-Sojo, J. Sun. Multi-layer Representation Learning for Medical
Concepts. KDD 2016: 1495-1504
differently from traditional classification models in which a [8] https://www.emc.com/collateral/analyst-reports/idc-extracting-value-
unique target is of interest, in Doctor AI the entire history of from-chaos-ar.pdf
the patient can be assessed. Deep Patient fed its neural network [9] E. Choi, M. Taha Bahadori, A. Schuetz, W. F. Stewart, J. Sun. Doctor AI:
Predicting Clinical Events via Recurrent Neural Networks. MLHC 2016:
with deidentified data from 700,000 EHRs and randomly 301-318.
combined them to make and test new variables for disease [10] R. Miotto, Li Li, Brian A. Kidd, Joel T. Dudley Deep Patient: An
risk. In order to describe the system we refer to Figure 6 Unsupervised Representation to Predict the Future of Patients from the
Electronic Health Records. www.nature.com/scientificreport
provided in [9].

View publication stats