BigTexts

A Framework for the Analysis of Electronic Health Record Narrative Texts based

on Big Data Technologies

Wilson Alzate Calder

, Alexandra Pomares Quimbaya

, Rafael A. Gonzalez

and Oscar Mauricio Mu

noz

Departamento de Ingenier

ıa de Sistemas, Pontiﬁcia Universidad Javeriana, Bogot

a, Colombia

Departamento de Medicina, Pontiﬁcia Universidad Javeriana, Hospital Universitario San Ignacio, Bogot

a, Colombia

Keywords:

Electronic Medical Record, Big Data, Natural Language Processing, Text Mining, Framework.

Abstract:

In the healthcare domain the analysis of Electronic Medical Records (EMR) may be classiﬁed as a Big Data

problem since it has the three fundamental characteristics: Volume, Variety and Speed. A major drawback is

that most of the information contained in medical records is narrative text, where natural language processing

and text mining are key technologies to enhance the utility of medical records for research, analysis and

decision support. Among the tasks performed for natural language processing, the most critical, in terms of

time consumption, are the pre-processing tasks that give some structure to the original non-structured text.

Studying existing research on the use of Big Data techniques in the healthcare domain reveals few practical

contributions, especially for EMR analysis. To ﬁll this gap, this paper presents BigTexts, a framework that

provides pre-built functionalities for the execution of pre-processing tasks over narrative texts contained in

EMR using Big Data techniques. BigTexts enables faster results on EMR narrative text analysis improving

decision making in healthcare.

1 INTRODUCTION

Nowadays, the amount of information stored in the

world is estimated at around of 1,200 exabytes (10

bytes), which come from multiple sources; for exam-

ple, Google processes over 24 petabytes (10

bytes)

of data per-day and Facebook gets more than 10

million photos every hour (Mayer-Schonberger and

Cukier, 2013). Given these large amounts of data

that are currently being generated, the paradigm of

Big Data has been proposed to describe the phe-

nomenon and suggest the need for organizations to

adopt innovative tools that can help to accurately de-

termine opportunities for improvement and creating

value (Moore et al., 2013). A Big Data problem is

one that complies with the three Vs: Volume (scale

of data), Variety (different forms of data) and Veloc-

ity (on data ﬂow analysis) (McAfee and Brynjolfsson,

2012). As in other domains, the healthcare domain

is generating increasingly large amounts of data, sig-

niﬁcantly within Electronic Medical Records (EMR).

Such data becomes an invaluable source for analy-

sis and research to address public health issues, im-

prove service quality and increase the coverage and

efﬁciency of health services (Sengupta, 2013).

Considering the large volume of data contained

in EMR, the variety going from structured to un-

structured data, and the speed which certain stud-

ies require, the analysis of EMR using computational

techniques may be classiﬁed as a Big Data problem

(Moore et al., 2013). In particular, the existence of

narrative text in ﬁelds such as nursing notes, treatment

plans, and lab tests, among others, generates a major

challenge that needs to be addressed natural language

processing technologies, text mining and Big Data.

To face this requirement from the healthcare do-

main, the aim of this paper is to present Big Texts,

a Big Data framework for the pre-processing of nar-

rative texts contained in EMRs. Big Texts provides

pre-built functions for the execution of pre-processing

tasks required to apply text mining techniques like

tokenization and Part-of-Speech recognition over the

text. In order to improve the performance of pre-

processing tasks, Big Texts uses big data technologies

like Map Reduce, Yarn and HDFS, and can be used

by running a standalone application or using its Java

Application Programming Interface (API).

In order to present Big Texts, this paper is or-

ganized as follows: Section 2 presents an overview

129

Alzate Calderón W., Pomares Quimbaya A., A. Gonzalez R. and Muñoz O..

BigTexts - A Framework for the Analysis of Electronic Health Record Narrative Texts based on Big Data Technologies.

DOI: 10.5220/0005434101290136

In Proceedings of the 1st International Conference on Information and Communication Technologies for Ageing Well and e-Health (ICT4AgeingWell-

2015), pages 129-136

ISBN: 978-989-758-102-1

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

of Big Data technologies and some related previous

work. Section 3 speciﬁes the characteristics and ar-

chitecture of Big Texts. Then Section 4 describes the

evaluation of the framework and its use for the analy-

sis of patient with a chronic disease. Finally, Section 5

presents conclusions and future work.

2 RELATED WORKS

Big data techniques have already been used by dif-

ferent projects. This section describes the most rele-

vant techniques related to Big Data; then, it presents

a set of related proposals for using them to improve

the analysis of data in the healthcare domain as well

as those created to enhance the execution of pre-

processing tasks and text mining techniques in any

context.

2.1 Big Data Technologies

Big Data is not a technology by itself, it is rather a

paradigm that has promoted the creation of several

products and frameworks; Table 1 presents a subset

of them including their classiﬁcation according to the

V property they are intended for.

Table 1: Big Data Technologies Classiﬁcation.

Technology Sub-technology Velocity Variety Volume

NoSQL X X

Analytics

Exploratory Data Analysis (EDA) X X

Conﬁrmatory Data Analysis (CDA) X X

Qualitative Data Analysis (QDA) X X X

MapReduce X X

Apache Hive X X

Apache Pig X X

MapReduce: is a computing style that can be

used to manage large-scale calculations in a way that

is hardware fault tolerant (Rajaraman and Ullman,

2012).

NoSQL: or Not Only SQL, is a data manage-

ment approach and database design that is useful for

very large datasets that are distributed (Sadalage and

Fowler, 2012).

Analytics: is the science of examining raw data

with the purpose of obtaining conclusions and convert

them to information (Maheshwari, 2015).

Hadoop: is a framework that allows distributed

processing of large sets of data across computer

clusters using simple programming models (White,

2012).

Apache Hive: provides a SQL-based dialect

called Hive Query Language (HiveQL) to query data

stored in a Hadoop cluster (Capriolo et al., 2012).

Apache Pig: provides an engine for a parallel data

ﬂow execution in Hadoop and a language called Pig

Latin for the expression of such data ﬂows (Gates,

2011).

2.2 Big Data Applied to the Health Care

Domain

Table 2 summarizes relevant projects related to the ap-

plication of Big Data technologies in the healthcare

domain. One of the main conclusion of this anal-

ysis is that they are mostly theoretical proposals of

desirable uses of Big Data technologies in that do-

main. Similarly, there are few tangible practices and

approaches of using this paradigm in products or pro-

totypes that make the expected beneﬁts a reality. Also

it was evident that one of the recurring factors in pa-

pers is the high amount of data required to analyze by

medical professionals and administrative personal to

make better decisions. Having in mind that one of the

major sources of information is the narrative text in

the EMRs, it is worth reviewing previous work that

use the Big Data paradigm to enhance the manage-

ment of textual data (Section 2.3).

2.3 Big Data Solutions to Improve Text

Analysis

The literature review of solutions that provide natu-

ral language processing and/or text mining techniques

indicates that there is an important platform called

GATECloud.net that provides several strategies to im-

prove response time in natural language processing,

among which has two options: the use of MapReduce

and the Cloud Computing technologies management.

From these two options the authors of these prod-

ucts recommend the second one, under a model PaaS,

since it is not necessary to rewrite the processing al-

gorithms to be executed. GATECloud.net is a Web

platform where scientists can conduct experiments on

text mining, and whose services are collected through

payment for use. It is important to note that this prod-

uct does not provide libraries (APIs) that can be used

from other systems (Tablan et al., 2012).

In addition to this proposal, there are other

projects that create customized solutions that use Big

Data technologies to analyze a speciﬁc set of nar-

rative texts. It is the case of the project presented

in (Das and Mohan Kumar, 2013) that proposes the

use of Big Data technologies and text mining tech-

niques to analyze public tweets. It uses technologies

such as HBase, Java, REST and Hadoop to enhance

the performance of the analysis. Even though these

kind of projects are important to mature the conjunc-

tion of Big Data and Text Mining, they do not pro-

vide open products or frameworks that can be used by

ICT4AgeingWell2015-InternationalConferenceonInformationandCommunicationTechnologiesforAgeingWelland

e-Health

130

Table 2: Comparison of Big Data Projects in the HealthCare Domain.

Project Technology Problem Solved Type Content

Use of Big Data and Cloud Computing for opera-

tional Health BI (Purkayastha and Braa, 2013)

Cloud Computing Bridging the digital divide for Medical In-

formation Systems in developing countries

Practical Text

Big Data in personalized health care (Chawla and

Davis, 2013)

Data mining Risk Proﬁle Custom diseases Theoretical

- Practical

Text

Big Data in Functional Echocardiography (Sen-

gupta, 2013)

Cloud computing, Robots,

Artiﬁcial Intelligence

To have more variables in the functional

echocardiography

Theoretical Images

Acquisition, compression, encryption and storage

of Big Data in Health (Brinkmann et al., 2009)

File processing Compressing data from electrophysiologi-

cal recordings

Practical Images

Big Data in medical devices (Meeker and Hong,

2014)

Sensors Reliability and availability of medical de-

vices

Theoretical Text

Development of an ecosystem of healthcare using

IPHIs (Liyanage et al., 2012)

Ontologies Ecosystem health interoperability for het-

erogeneous systems

Theoretical Text

Big Data to assess the impact of medical pro-

grams (Fox, 2011)

Analytics Allocate and keep real patients Theoretical Text

other projects, which is our need to analyze medical

records.

As is evident, although there are approaches that

try to unify the paradigm of Big Data to text process-

ing, they are not still mature enough or do not provide

APIs through which existing systems can run tasks

(such as pre-processing) in large amounts of text, pro-

viding a great opportunity to attacked by BigTexts

(Section 3), the proposal of this paper.

3 BIGTEXTS

BigTexts is a framework that allows the execution of

pre-processing tasks over narrative texts required to

apply text mining techniques. It was developed using

Big Data technologies, which allows it to improve the

performance when it is necessary to analyze large vol-

umes of medical records and texts contained in them.

BigTexts allows the pre-processing of hundreds, thou-

sands or millions of narrative texts when a research

required the analysis at institutional, regional or na-

tional level. BigTexts provides pre-built functional-

ities that can be used as an independent application

(with a graphical user interface - GUI) or as a library

(API) from an existing application.

Figure 1 shows highlighted the BigData technolo-

gies used in BigTexts. BigTexts uses Apache Pig

at the data access level, MapReduce for distributed

processing, YARN as a resource manager, HDFS for

distributed storage and Linux Ubuntu as the operat-

ing system. BigTexts, takes the EMR texts and par-

tition them in the Hadoop cluster using HDFS. Us-

ing MapReduce to execute the pre-processing tasks

in parallel, in order to ﬁnally obtain the set of pre-

processed documents.

As it is shown in Figure 2, typically the analysis

of EMRs are executed sequentially (As-Is). This way

of running generates a bottle neck when the number

of EMR is large. With BigTexts the execution is par-

allelized across the set of machines available in the

Figure 1: BigTexts in the Reference Architecture.

cluster where BigTexts is running. BigTexts takes the

EMR texts and partitions them in the Hadoop cluster

using HDFS. Using MapReduce it executes the pre-

processing tasks in parallel. Finally, it obtains the set

of pre-processed texts.

The following sections explain the functionalities

provided by BigTexts and its architecture.

3.1 BigTexts Tasks

As it was mentioned, BigTexts provides pre-

processing tasks required to give some structure to

the originally unstructured texts. These tasks allow

for example the division of the text in tokens, to iden-

tify sentences, to recognize the role of each word in

a sentence, to recognize patterns in the text, etc. The

following list explains the current tasks provided by

BigTexts, which are based on the Stanford CoreNLP

library (The Stanford NLP Group, 2014b) and the

Weka library (The University of Waikato, 2014). It

is important to notice that these tasks can be extended

according to the requirements of the analysis.

• Identiﬁcation of co-reference: It identiﬁes when

two or more expressions in a text refer to the

BigTexts-AFrameworkfortheAnalysisofElectronicHealthRecordNarrativeTextsbasedonBigDataTechnologies

131

Figure 2: Current Situation vs. BigTexts Situation.

same person or thing (The Stanford NLP Group,

2014a).

• Lemmatisation Lovins: It reduces a word to its

lemma, root or canonical form using the Lovins

Stemmer algorithm (Lancaster University, 2014).

• Named Entity Recognition: It locates and gives

a label to some elements in text taking into ac-

count pre-deﬁned categories such as the names of

persons, organizations, locations, dates, monetary

values, etc (The Stanford NLP Group, 2014d).

• Part of Speech: It recognizes and gives a label in-

dicating the grammar role a word plays in a sen-

tence (The Stanford NLP Group, 2014c).

• Named Entity Recognition based on regular ex-

pressions (NE Transducer): It locates and gives a

label to some elements in text taking into account

rules described using regular expressions.

• Lemmatisation Snowball: It reduces a word to its

lemma, root or canonical form using the snowball

stemmer algorithm.

• Partition text by phrases: It splits a text into its

sentences using a set of conﬁgurable terminators

like ., ?.

• Tokenization: It breaks down a text into tokens

using a set of conﬁgurable separators like space,

punctuation tokens, among others.

• Switch a text to upper case.

The implementation of the pre-processing tasks

was made using Apache Pig, particularly the option

for creating User Deﬁned Functions (UDFs). Each

pre processing tasks is represented by a Java class

that inherits from org.apache.pig.EvalFunc and pack-

aged using Apache Maven. After that the resulting

jar is added to the Pig pipeline and the UDF is called

inside of Pig instructions. As mentioned before, the

pre-processing tasks use two NLP libraries, Stanford

CoreNLP and Weka.

The addition of a new pre-processing task is de-

signed to be a simple and ﬂexible process. To do this,

a new class is created in which the functionality is im-

plemented and added to the preprocessing tasks cata-

log

An XML ﬁle with all the pre-processing tasks and their

parameters.

Figure 3: BigTexts Use Cases.

Each one of the previous tasks can be executed

using an API provided by BigTexts or directly using

its graphical user interface. Figure 3 describe the use

cases of BigTexts:

1. CU001-Upload documents: The actor can load a

set of documents for further processing.

2. CU002-Select preprocessing tasks: The actor can

select the pre-processing tasks he/she wants to ex-

ecute in BigTexts.

3. CU003-Establish execution order: The actor can

give the order in which the selected tasks must be

executed and conﬁgured on the CU002.

4. CU004-Select destination result: The actor

chooses the way the results are going to be deliv-

ered, with two options: FTP or HDFS directory.

5. CU005-Execute preprocessing tasks: The actor

can launch the execution of the pre-processing

task list over the uploaded documents.

3.2 BigText Process and Architecture

Figure 4 shows an overview of how BigTexts interacts

with the BigData technology. As it can be seen it is

composed by a Client component (BigTexts Client)

and a Server component(BigTexts).

The Client offers a graphical user interface

through which the user can select the ﬁles to be pro-

cessed, the pre-processing tasks to run (and their set-

ICT4AgeingWell2015-InternationalConferenceonInformationandCommunicationTechnologiesforAgeingWelland

e-Health

132

tings) and the delivery method. This client applica-

tion can also be used by an external application as a

library. Once a user sends the ﬁles to the server using

FTP and indicates the tasks he/she wanted to execute,

the client component sends an XML format message

to the server queue.

The BigTexts server component connects to the

server queue (ActiveMQ) and gets the queued mes-

sage, transforming it from XML (using the library

JAXB

) to objects. Subsequently, the converted mes-

sage to object is executed in the Hadoop cluster

by an Apache Pig script that uses UDFs (User De-

ﬁned Functions) with the pre-processing tasks. The

Hadoop cluster is composed of a main machine (the

master), which manages both storage and parallel pro-

cessing, and a number of secondary machines (slaves)

that stores data blocks and process them in parallel.

Finally, a set of pre-processed documents are deliv-

ered in the HDFS or in the FTP directory, according

to what the user has selected.

Figure 4: BigTexts.

The internal architecture of BigTexts Server

presented in Figure 5.

1. pig-client-0.13.0.jar: It is the library that allows

the connection with the installation of Pig.

2. bigtexts-util-1.0.jar: BigTexts utility classes.

3. log4j-1.2.17.jar: The application log manager.

4. hadoop-client-2.5.0.jar: Library that allows inter-

action with HDFS and Hadoop.

5. bigtexts-dto-1.0-jar: Package with the classes that

allow the communication between the BigTexts

client and server.

6. automaton-1.11-8.jar: Finite state automata li-

brary, used by the Pig client.

7. eclipselink-2.3.2.jar: Java Persistence API imple-

mentation created by The Eclipse Foundation

Java Architecture for XML Binding

bigtexts-1.0.jar: Server application, is responsible for ex-

ecuting the pre-processing tasks into the BigData infras-

tructure.

Figure 5: BigTexts Physical View.

8. bigtexts-udfs-1.0.jar: Component with the imple-

mentations of the pre-processing tasks.

(a) libstemmer-1.0.jar: Library with the Snowball

Stemmer implementation.

(b) stanford-corenlp-3.4.1.jar: Natural Language

Processing library.

mentation of Lovins Stemmer.

4 CASE STUDY AND

VALIDATION

In order to validate the functionality and the perfor-

mance of BigTexts, the framework was used for the

analysis of EMRs contained in a Hospital EMR Sys-

tem. The goal of the analysis was to identify a set of

patients with speciﬁc characteristics required by med-

ical researchers for a retrospective cohort study. This

section presents ﬁrst the scenario in which the vali-

dation of the system was performed (Section 4.1) and

then its results (Section 4.2).

4.1 Validation Scenario

The validation scenario requires searching, from a set

of EMRs, those records that match different charac-

teristics required to include them in a study on Heart

Failure. The input for the analysis of each record in-

cludes terms related to the diagnosis, personal and

BigTexts-AFrameworkfortheAnalysisofElectronicHealthRecordNarrativeTextsbasedonBigDataTechnologies

133

family antecedents, lab exams, medications, symp-

toms, treatment plans complications and outcomes.

The intention of the case study was to ﬁnd the set of

EMRs that include most of the characteristics deﬁned

by the group of researchers to include them in a ret-

rospective cohort study around this chronic disease.

BigTexts was used for this purpose. It was conﬁg-

ured for the pre-processing of 10.000 narrative texts

provided by the hospital. In order to assure the conﬁ-

dentiality of the patients the records were previously

anonymized and provided in 4 ﬁles. In addition, the

Named Entity Recognition task of BigTexts was con-

ﬁgured for the identiﬁcation of terms related to the

searched diagnosis like Cardiac Failure, Cardiopathy,

chronic heart failure, CHF, among others;related to

lab exams like bnp, echocardiogram; related to symp-

toms like Paroxysmal Nocturnal Dyspnea, and related

to complications as death.

From the technical point of view, the machines

used for the conﬁguration of the Hadoop cluster (Ta-

ble 3) were not high speciﬁcation computers (server

type), but rather common computers in order to simu-

late real hospital environments, where was not pos-

sible to have dedicated servers. A main machine

(master) was used and a set of secondary machines

(slaves). The client application was installed on the

master machine as well as the BigTexts server. All

computers were installed with the Ubuntu operating

system 14.0.LTS. The machines were connected in a

local area network (LAN) establishing static IP ad-

dresses. All machines had Intel processors.

The size of the ﬁles provided was the following:

1.945.886 bytes (1.945 Mb), 1.423.803 bytes (1.423

Mb), 437.643 bytes (437.643 Kb) and 761 bytes. The

validation consisted of the execution of ﬁve iterations

per ﬁle in each of the following pre-processing tasks:

• RegexNamedRecognition: Given the list of terms

related to the diagnosis and the lab exams with a

label for each of them, the system identiﬁed using

regular expressions if a word was in the token list

and set the corresponding label.

• Tokenizer: Partition ﬁles into tokens.

• Tokenizer + POS-Tagger: Two preprocessing

tasks were performed. The ﬁrst one broke up the

ﬁle into tokens and then applied Part of Speech

task identifying whether each word was a verb, an

adjective, an adverb, etc.

• Tokenizer + SnowballStemmer: It breaks up the

ﬁle into tokens and then identiﬁed the root of each

one of them.

4.2 Perfomance Results

The analysis of the data, using ﬁgures of lines to iden-

tify trends is shown below. For a clear display on the

graph the ﬁle size was divided for 10.000, having all

of them in the same scale. There was used a combi-

nation of N, M and L machines, where N < M < L.

Finding that the Regex Entity Recognition task

(Figure 6) execution with the 1,945,886 bytes ﬁle and

N machines obtained an average time of 109.13 sec-

onds, for M machines of 92.78 seconds and for L

machines of 82.25 seconds. For the 1.423.803 bytes

ﬁle, the times were 104.07, 98.33 and 80.45 sec-

onds. Likewise, the ﬁle of 437,643 bytes had times

of 70.54,65.13 and 58.71 and the ﬁle of 761 bytes had

times of 61.57, 54 and 51 seconds for N, M and L ac-

tive machines in the cluster. As it can be seen, there

is a downward trend on time when the number of ma-

chines is increased.

Figure 6: Validation - RegexNER.

For the tokenization task (Figure 7) for the ﬁle

of 1.945.886 bytes, times obtained were 50.22, 46.63

and 36 seconds. For the ﬁle of 1.423.803, time av-

erage was 56.25, 41.33 and 33.6 seconds. Similarly,

for the 437.643 ﬁles and 761 bytes, execution times

improved when the number of machines in the cluster

was increased.

Regarding the tokenization task added to the Part

of Speech (Figure 8) task, for the 1.945.886 bytes

ﬁles, times were 181, 113.75 and 85.67 seconds; for

the 1.423.803 bytes the times were 118.33, 96.67 and

76.33 seconds; and for the 761 bytes the times were

69.67 , 65 and 52.5 seconds. These results show an

improvement when the number of available machines

in the cluster were increased. However, in the case of

the 437.643 bytes ﬁle there was a not an improvement

in time when the number of machines was increased

slightly, but a substantial improvement was presented

when the number of machines was changed radically.

For the tokenization task added to Stemming using

ICT4AgeingWell2015-InternationalConferenceonInformationandCommunicationTechnologiesforAgeingWelland

e-Health

134

Table 3: Cluster Machines.

Name Hostname IP Interna Brand Memory Hard drive Processor

bigtexts-1 master 192.168.0.101 HP 3,8 GiB 155,3 GB Intel Core 2 Duo CPU E7200 2.53GHz x 2

bigtexts-2 slave-2 192.168.0.102 HP 1.9 GiB 155.3 GB Intel Core 2 Duo CPU E7200 2.53GHz x 2

bigtexts-3 slave-3 192.168.0.103 Lenovo 1.9 GiB 76.5 GB Intel Core 2 CPU 4400 2.00GHz x 2

bigtexts-4 slave-4 192.168.0.104 HP 1.9 GiB 155.3 GB Intel Core 2 Duo CPU E4600 2.40GHz x 2

bigtexts-5 slave-5 192.168.0.105 HP 1.9 GiB 155.3 GB Intel Core 2 Duo CPU E7200 2.53GHz x 2

Snowball algorithm (Figure 10) the following times

for N, M and L machines were obtained, respectively:

• 1.945.886 bytes: 69.15, 53.86 and 49 seconds

• 1.423.803 bytes: 47.05, 45.4 and 43.71 seconds

• 437.643 bytes: 46.46, 43.6 and 40.67 seconds

• 761 bytes: 50.21, 47.42 and 42.25 seconds

Figure 7: Validation - Tokenizer.

Figure 8: Validation - Tokenizer + POS.

Allowing identifying a downward trend as they

were adding available machines to the cluster as a

possible option.

As it can be seen the execution time for the same

ﬁle in a same task, differs depending on the number

of available slaves. It improves as new slaves were

added to the cluster. Besides, it can be seen that the

improvement is much more noticeable when complex

tasks need to be executed like RegexNER and Part Of

Figure 9: Validation - Tokenizer + Snowball.

Speech. Additionally, the larger is the ﬁle, the greater

the processing speed, for example, for the 761 bytes

ﬁle the times for different scenarios (N, M and L ma-

chines) are very similar.

From the point of view of the requirements of the

medical researchers BigTexts allowed the identiﬁca-

tion of the set of EMRs that include all or part of the

terms required for the analysis. The result allowed re-

ducing the time previously required to select records

and analyze which of them actually met the charac-

teristics desired for the study.

5 CONCLUSIONS AND FUTURE

WORK

Shortening large EMR data analysis helps clinical and

administrative data management in the health domain

with both decision making and quality improvement

as beneﬁts.

BigTexts demonstrated its utility to improve efﬁ-

ciency in narrative text analysis of electronic medi-

cal records, this will not only allow performing more

analysis projects in the health sector but also al-

low generating new products that can be constructed

based on BigTexts.

Additionally, BigTexts showed that the addition of

an abstraction layer to the entire ecosystem of tech-

nology of Big Data and text mining avoids installa-

tion, conﬁguration and integration efforts of the Big

Data components that, although powerful, are limited

BigTexts-AFrameworkfortheAnalysisofElectronicHealthRecordNarrativeTextsbasedonBigDataTechnologies

135

in terms of accurate and reliable documentation.

The validation results let us to conclude that it

is worth to use the entire infrastructure of Big Data

when the analysis include large ﬁles or complex tasks,

because it was found that only in those situations the

addition of nodes into the cluster markedly improves

performance.

As future work it is proposed to use cloud comput-

ing to address the problems encountered in the present

project, to evaluate the changes in development ef-

fort, and to promote scalability of BigTexts. Addi-

tionally, another ﬁeld of work is the development of

more UDFs to increase the catalogue of BigTexts pre-

processing tasks.

REFERENCES

Brinkmann, B. H., Bower, M. R., Stengel, K. A., Wor-

rell, G. A., and Stead, M. (2009). Large-scale elec-

trophysiology: Acquisition, compression, encryption,

and storage of big data. Journal of Neuroscience

Methods, 180(1):185–192.

Capriolo, E., Wampler, D., and Rutherglen, J. (2012). Pro-

gramming Hive. O’Reilly Media, 1 edition edition.

Chawla, N. V. and Davis, D. A. (2013). Bringing big

data to personalized healthcare: A patient-centered

framework. Journal of General Internal Medicine,

28(S3):660–665.

Das, T. and Mohan Kumar, P. (2013). Big data analytics:

A framework for unstructured data analysis. School

of Information Technology and Engineering, VIT Uni-

versity.

Fox, B. (2011). Using big data for big impact. leveraging

data and analytics provides the foundation for rethink-

ing how to impact patient behavior. Health manage-

ment technology, 32(11):16.

Gates, A. (2011). Programming Pig. O’Reilly Media, 1

edition edition.

Lancaster University (2014). What

is stemming? Retrieved from

http://www.comp.lancs.ac.uk/computing/research

stemming/general/.

Liyanage, H., Liaw, S.-T., and de Lusignan, S. (2012). Ac-

celerating the development of an information ecosys-

tem in health care, by stimulating the growth of safe

intermediate processing of health information (IPHI).

Informatics in primary care, 20(2):81–86.

Maheshwari, A. (2015). Data Analytics Made Accessible.

Mayer-Schonberger, V. and Cukier, K. (2013). Big Data: A

Revolution That Will Transform How We Live, Work,

and Think. Houghton Mifﬂin Harcourt, Boston.

McAfee, A. and Brynjolfsson, E. (2012). Big data: The

management revolution. Harvard business review,

90(10):p60–68.

Meeker, W. Q. and Hong, Y. (2014). Reliability meets big

data: Opportunities and challenges. Quality Engineer-

ing, 26(1):102–116.

Moore, K. D., Eyestone, K., and Coddington, D. C. (2013).

The big deal about big data. Healthcare ﬁnancial

management: journal of the Healthcare Financial

Management Association, 67(8):60–66, 68. PMID:

23957187.

Purkayastha, S. and Braa, J. (2013). Big data analytics for

developing countries – using the cloud for operational

BI in health. The Electronic Journal of Information

Systems in Developing Countries, 59(0).

Rajaraman, A. and Ullman, J. D. (2012). Mining of mas-

sive datasets. Cambridge University Press, New York,

N.Y.; Cambridge.

Sadalage, P. J. and Fowler, M. (2012). NoSQL Distilled:

A Brief Guide to the Emerging World of Polyglot Per-

sistence. Addison-Wesley Professional, Upper Saddle

River, NJ, 1 edition edition.

Sengupta, P. P. (2013). Intelligent platforms for dis-

ease assessment. JACC: Cardiovascular Imaging,

6(11):1206–1211.

Tablan, V., Roberts, I., Cunningham, H., and Bontcheva,

K. (2012). GATECloud.net: a platform for large-

scale, open-source text processing on the cloud.

Philosophical Transactions of the Royal Society A:

Mathematical, Physical and Engineering Sciences,

371(1983):20120071–20120071.

The Stanford NLP Group (2014a). Coref-

erence resolution. Retrieved from

http://nlp.stanford.edu/projects/coref.shtml.

The Stanford NLP Group (2014b). Stan-

ford CoreNLP. Retrieved from

http://nlp.stanford.edu/software/corenlp.shtml.

The Stanford NLP Group (2014c). Stanford log-

linear part-of-speech tagger. Retrieved from

http://nlp.stanford.edu/software/tagger.shtml.

The Stanford NLP Group (2014d). Stanford named

entity recognizer (NER). Retrieved from

http://nlp.stanford.edu/software/CRF-NER.shtml.

The University of Waikato (2014). Weka 3

- data mining with open source machine

learning software in java. Retrieved from

http://www.cs.waikato.ac.nz/ml/weka/.

White, T. (2012). Hadoop: The Deﬁnitive Guide. Yahoo

Press, Beijing, third edition edition edition.

ICT4AgeingWell2015-InternationalConferenceonInformationandCommunicationTechnologiesforAgeingWelland

e-Health

136