=Paper=
{{Paper
|id=Vol-2022/paper38
|storemode=property
|title=
Integrating Data Analysis Tools for Better Treatment of Diabetic Patients
|pdfUrl=https://ceur-ws.org/Vol-2022/paper38.pdf
|volume=Vol-2022
|authors=Svetla Boytcheva,Galia Angelova,Zhivko Angelov,Dimitar Tcharaktchiev
|dblpUrl=https://dblp.org/rec/conf/rcdl/BoytchevaAAT17a
}}
==
Integrating Data Analysis Tools for Better Treatment of Diabetic Patients
==
https://ceur-ws.org/Vol-2022/paper38.pdf
Integrating Data Analysis Tools for Better Treatment of
Diabetic Patients
Svetla Boytcheva1, Galia Angelova1, Zhivko Angelov2, Dimitar Tcharaktchiev3
1
Institute of Information and Communication Technologies, Bulgarian Academy of Sciences,
Sofia, Bulgaria
2
Adiss Lab Ltd., Sofia, Bulgaria
3
Medical University Sofia, University Specialized Hospital for Active Treatment of
Endocrinology, Sofia, Bulgaria
svetla.boytcheva@gmail.com, galia@lml.bas.bg, angelov@adiss-bg.com, dimitardt@gmail.com
Abstract. This paper presents the construction and usage of an anonymous Diabetes Register for
patients in Bulgaria. The Register is generated automatically from outpatient records submitted to the
Bulgarian National Health Insurance Fund in 2010-2014 and continuously updated using outpatient records
for 2015-2016. The construction relies on advanced automatic analysis of free text information as well as on
Business Analytics technologies for storing, maintaining, searching, querying and analyzing data. Original
frequent pattern mining algorithms enable to find patterns and sequences taking into account temporal
information. The paper discussed the software environment as well as experiments in frequent pattern mining
that enable knowledge discovery in the very large repository underlying the Register (currently 262 million
pseudonymized outpatient records submitted to the Bulgarian National Health Insurance Fund in 2010-2016
for more than 5 mln citizens yearly). The claim is that the synergy of modern analytics tools transforms a
static archive of clinical patient records to a sophisticated software environment for knowledge discovery and
prediction.
Keywords: synergy of data management, data mining and text mining tools; clinical data; frequent
pattern mining; data analytics; natural language processing; knowledge discovery
In this paper we present the integration of various
1 Introduction ICT tools for automatic generation of a Diabetes Register
for Bulgarian patients. The huge amount of clinical data,
Medicine is known as a Data Intensive Domain: due to underpinning a repository of Outpatient Records (ORs),
the recent penetration of the Information and enabled to construct interfaces that support both
Communication Technologies (ICT) in all areas of our monitoring functionalities (oriented to the health
society, a rapidly increasing amount of medical data is management authorities) and research-oriented
produced by the healthcare sector, on the one hand, and functionalities for knowledge discovery. The monitoring
by biomedical research on the other hand. In the functionalities are based on business analytics while
healthcare sector, ICT applications support health research tools use data mining and pattern search. The
diagnostics, development and maintenance of medical software environment includes also components for
Electronic Health Records, telemedicine and telecare, automatic analysis of free texts in Bulgarian. These
patient administration, almost all aspects of healthcare components facilitate the Register generation and its
management and healthcare delivery as well as medical update because they deliver values of clinical tests and
education and training. In biomedical research, progress lab data which are described as unstructured text only.
in deeper understanding of medical phenomena is sought This paper is structured as follows. Section 2
by construction of big data models: e.g. virtual overview related work in several areas that are relevant
physiological human, models of brain, in computational to the subject: Diabetes registers, Natural Language
genetics and so on. Public access to health information is Processing (NLP) for clinical narratives, Business
changing the relationship between the patients and the Intelligence (BI) and analytics, Frequent Pattern Mining
health institutions that are responsible for care delivery. (FPM). Section 3 presents the experimental study context
The monitoring and control function of patient and summarizes the developments during the last 3-4
organizations is facilitated by the modern ICT tools as years (because the Register was built iteratively). Section
well. Today we are still in an early phase of a long-term 4 presents relevant achievements in automatic analysis of
technological and social shift that will be implied by clinical narratives in Bulgarian language. Section 5
advancing further the ICT fundamentals and tools. discusses recent algorithms for frequent pattern mining
and presents experiments related to knowledge discovery
Proceedings of the XIX International Conference in the Register repository. Section 6 contains the
“Data Analytics and Management in Data Intensive conclusion and plans for future work.
Domains” (DAMDID/RCDL’2017), Moscow, Russia,
October 10-13, 2017
230
2 Related Work sensitivity of the recognition. Despite the NLP
limitations, the conclusion is that NLP engines are
2.1 Diabetic Registers powerful components ready for integration in medical
data mining and – due to improvements expected in the
There are several nation-wide Diabetes Registers in the
future, e.g. more accurate mappings of terms to medical
world, e.g. in Denmark [1], Sweden [2], Norway [3].
concepts – the importance of NLP as a valuable
Registers explicate the number of patients who are
supporting technology will grow.
diagnosed with Diabetes and provide good monitoring
Here we consider NLP for Bulgarian clinical text. No
and control. Constructing registers is expensive and
comprehensive resources exist for Bulgarian medical
burdening the patients as well as the medical experts with
language; the International Classification of Diseases
additional administrative work. Furthermore, in some
ICD-10 is the only terminological resource which is
countries chronic disease management is not recognized
available in electronic format. Our experience shows that
as a part of general medical practice. As for the
within 2-3 years one can achieve good performance in
construction, most medical experts agree that Registers
separate extraction tasks. We apply software prototypes
are a must since Diabetes is a chronic disease with
developed some years ago that are gradually improved.
significant social consequences. Electronic patient
The most useful tools are a drug extractor (it finds in the
registration systems are proposed like the one in Ireland
free text the drug name, dosage, frequency and route of
[4] (but it is not implemented yet). It is interesting to
admission [13]) as well as an extractor of numeric values
mention that in Sweden, during the Diabetic Register
of lab data and clinical tests [14].
development phase 2001-2005, the registration rate of
patients gradually increased and reached 75% which in 2.3 Big Data, Business Intelligence Tools
2010 still remains stable and is one of the highest in the
country [5]. Thus infrastructure construction is a critical Big Data usually designates a massive volume of
issue but data collection and update are further problems structured and unstructured data, too large or too
that can be solved only by persistency and diligence. dynamic to be processed by traditional software tools and
techniques. The popular "3Vs" features of Big Data were
2.2 NLP of Clinical Narratives first introduced by Gartner (previously META group):
"high Volume, high Velocity, and/or high Variety" [15].
Usually automatic analysis of clinical narratives is
Wikipedia is an example for big data consisting of
implemented partially: only fragments of the text are
unstructured texts, images and hyperlinks. Big data
considered. The phrases, selected as “interesting”, are
analytics is the process of collecting, organizing and
typically picked up due to the presence of a word or an
analyzing big data to discover useful information.
entity which are considered “significant”. This approach
Business Intelligence tools analyze big data of
for shallow analysis is called “Information Extraction”
enterprises in order to provide historic, present and
(IE). IE from clinical texts matures only recently but its
predicted views to the business processes. Predictive
accuracy gradually improves and often exceeds 90% [6].
analytics for establishment of trends is the preferred
The review [6] stated in 2008 that “current applications
functionality in contrast to databases that deal with data
are rarely applied outside of the laboratories they have
items and extract subsets of data values. Visualization is
been developed in, mostly because of scalability and
an important feature of BI tools because they show
generalizability issues”. Today, however, this is valid for
generalizations and tendencies in one screen [16].
languages other than English because, with the active
Another necessary feature is the speed of processing
contribution of numerous research groups in the USA,
since big data often appear in real time.
NLP for English clinical narratives has much better
In our project we use a BITool which stores data in
performance at present. Comprehensive language
n-dimensional cubes and explores multi-dimensional
resources exist for English, such as UMLS [7] as well as
data i.e. hyperplanes [17]. The user can split the dataset
tools like KnowledgeMap Concept Identifier [8] which
into groups of objects with similar features. If temporal
processes clinical notes and returns CUIs (Concept
dimension is included the user can track changes of
Unique Identifiers) for the recognized UMLS terms.
object characteristics over time by animation. BITool
Another important tool is the public NegEx system
enables the discovery of similar situations over time
which identifies and interprets negations in English
when a search pattern is specified for a particular period.
clinical texts [9, 10]. We also mention the open-source
cTAKES1 (clinical Text Analysis and Knowledge 2.4 Frequent Pattern Mining
Extraction System) and the Health Information Text
Extraction (HITEx) system [11]. A recent study [12] There are two principal tasks in pattern search: frequent
enumerates the advantages to incorporate NLP for pattern mining (FPM) where the events (objects) are
English in medical systems: it systematically links considered as unordered sets, and frequent sequence
several terms to a concept using databases that mining (FSM). Approaches for solving the FPM task
standardize health terminologies; avoids manual work vary from the naïve BruteForce and Apriori algorithms,
for searching term variations; increases the number of where the search space is organized as a prefix tree, to
patients in the considered cohorts and thus increases the Eclat algorithm that uses tidsets directly for support
1
http://ctakes.apache.org/
231
computation by processing prefix equivalence classes they are extracted automatically from four XML fields:
[18]. Most FPM and FSM methods do not consider (i) Anamnesis: summarizes case history, previous
contextual information about extracted patterns. They treatments, often family history, risk factors; (ii) Status:
usually build a (huge) prefix tree. Most FPM algorithms summary of patient state, height, weight, BMI, blood
generate all possible frequent patterns (FPs). pressure etc.; (iii) Clinical tests: values of clinical
Summarized information for data relations can be examinations and lab data listed in arbitrary order; (iv)
extracted as maximal frequent itemsets (MFI) in order to Prescribed treatment: codes of drugs reimbursed by
reduce redundancy and decrease significantly the NHIF, free text descriptions of other drugs. Integration
number of FPs for post-analysis. All classic algorithms of large scale text analysis is a real novelty in this field.
for FPM can be modified for MFI search.
We have proposed a novel algorithm for mining sets 3.2 Analytics Using BITool
of events in order to identify strong co-occurrence of Today the system BITool supports the Diabetes Register
patterns [19]. It is a cascade data mining approach for at the University Specialized Hospital for Active
FPM enriched with context information which aims at Treatment of Endocrinology ″Acad. Ivan Penchev″,
the discovery of complex relations between medical Medical University – Sofia (this Hospital was authorized
events with respective timestamps. Experiments with by the Bulgarian Ministry of Health to host the Register
this approach are presented in Section 5 to illustrate the of diabetic patients in Bulgaria). BITool’s functionalities
functionality of the Diabetes Register as a research tool. enable the monitoring of significant indicators like
glycated hemoglobin (HbA1c) and blood glucose values.
3 Experimental Study In this way the Register achieves its objective: to provide
an adequate monitoring strategy for diabetic patients and
3.1 Principal Objective to improve the healthcare and quality of life for the
A pseudonymized Register of diabetic patients was patients and their families. Two examples illustrate the
generated in 2015 from the Outpatient Records, collected services. Figure 1 shows the number of diabetic patients
by the Bulgarian National Health Insurance Fund in the dimensions age-gender (at certain moment). Here
(NHIF), in compliance with all legal requirements for BITool operates on the structured information from the
safety and data protection [20]. The usual patient NHIF archive: patient pseudonym, age and gender.
registration process was kept without burdening the Further statistics of this kind might concern explorations
medical experts with additional paper work. NHIF is the of diabetic patients per region code, types of diabetes and
only obligatory Insurance Fund in Bulgaria so we note diabetes complications, per GPs, per types of medication,
that working with ORs ensures 100% registration of all according to frequency of visits etc.
patients who contacted the healthcare system at all
(however there are Bulgarian citizens who are not
insured and some others who have ORs but are not
properly diagnosed with Diabetes). The data repository,
underpinning the Register, currently contains more than
262 mln pseudonymised ORs submitted to the NHIF in
2010-2016 for more than 7.3 mln Bulgarian citizens
(more than 5 mln yearly), including 483,836 diabetic
patients. In Bulgaria ORs are produced by General
Practitioners (GPs) and Specialists from Ambulatory
Care whenever they contact patients. Despite the primary
accounting purpose ORs summarize sufficiently the case
Figure 1 Number of diabetic patients grouped by age
and motivate the requested reimbursement. They are
semi-structured files with predefined XML-format.
Many indicators in the Register copy the structured data
submitted to NHIF in ORs: (i) date and time of the visit;
(ii) pseudonymized personal data, age, gender; (iii)
pseudonymised visit-related information; (iv) diagnoses
in ICD-10; (v) NHIF drug codes for medications that are
reimbursed; (vi) a code if the patient needs special
monitoring; (vii) a code concerning the need for
hospitalization; (viii) several codes for planned
consultations, lab tests and medical imaging.
ORs contain also important values presented in free
text fields: glycated haemoglobin (HbA1c), body mass
index (BMI), weight, blood glucose and blood pressure Figure 2 Reduction of HbA1c levels after application
etc. These values are essential for a Diabetic Register so of incretin2 based drugs
2
https://www.drugs.com/drug-class/incretin-mimetics.html
232
Figure 2 explores the tendency in the development of 5 Research in Frequent Pattern Mining
treatment. It displays the number of patients who had
changes in the HbA1c levels within the interval [-5,5] 5.1 Contextual Information
units for certain period of time. For most patients the
Most FSM and FPM approaches do not use contextual
HbA1c level decreased by 1 unit. The HbA1c levels are
information about extracted patterns. These algorithms
extracted from the free text of ORs for the corresponding
extract general templates but do not answer the major
patients with timestamp.
question whether they are influenced in some way by the
Finally we show the Register interface during the
context and whether they are valid in various aspects.
process of exploring the collection of ORs (Figure 3).
Existing methods which search for patterns using
The names and personal identifiers of patients and GPs
contextual information are based on attributes that are
are replaced by pseudonyms; only the name of the
organized into hierarchical structures and on attributes’
city/village remains in the address field.
generalizations and specializations.
Context information is organized as attributes of
itemsets and tidsets. Attributes may have different
organization - structured or unstructured. This enables to
explore the context-dependent templates. Rabatel et al.
[22] propose an approach in marketing domain taking
into account not only the transactions that have been
made but also various attributes associated with
customers like age, gender etc. Attributes have a
hierarchical structure (𝐻(𝐴𝑔𝑒), 𝐻(𝐺𝑒𝑛𝑑𝑒𝑟)) and
explore patterns at different levels of attributes
abstraction – lattice 𝐻 (Figure 4). Traditional methods
Figure 3 Exploring outpatient records in the Register consider only the top level [∗,∗] - for any age and
regardless of gender, i.e. without attributes. Rabatel et al.
4 NLP for Bulgarian Clinical Narratives designed the algorithm Gespan and made experiments
with about 100,000 product descriptions from
Design and implementation of software for automatic amazon.com.
extraction of patient-related entities from a Big Data
collection is a quite challenging task. One needs to scale [*,*] H
up existing research prototypes to process millions of H(age) *
H(gender)
patient records, coping with noisy and missing data, and young old
[y,*] [o,*] [*,m] [*,f]
*
still providing reliable results. Some numeric entities (y) (o)
refer to key risk factors for development of Diabetes male
(m)
female
(f)
[y,m] [y,f] [o,m] [o,f]
Mellitus (levels of glycated hemoglobin HbA1c and
blood glucose) and cardio-vascular diseases (high blood Figure 4 Structuring attributes in marketing domain
pressure). Unfortunately in the Bulgarian clinical Ziembiński [23] proposes a new approach for
practice these values are usually documented in free text extracting small contextual models from smaller
paragraphs, presented in a huge variety of formats, so collections of data that later are summarized in
their automatic identification is difficult. We note that generalized models using information from contextual
according to some studies, today more than 80% of the models with common information. This approach applies
patient-related clinical information is stored as free text a metrics for measuring distance of context models. All
in the Electronic Health Record systems. values for similarity assessment are normalized in the
In [21] we proposed a hybrid method for automatic range between 0.0 and 1.0. Attribute values are
generation of grammar rules for IE from clinical data. considered identical if the similarity function returns 1.0.
The experiments were made and evaluated over In the opposite case the result is 0.0. This approach
approximately 9.5 million of ORs. Here we cite only the allows extracting patterns for data that would otherwise
evaluation of blood pressure extraction from the ORs of have to be dropped out of the templates because of its
about 1,800,000 patients with arterial hypertension for 3 dispersion and low frequency.
year period: all available values are about 38.3 million
and the extraction was performed with precision 92% 5.2 Experimental Setup
and recall 98%. The variety of recording formats and
explanations written by thousands of medical We apply a retrospective analysis for patients from the
professionals require constant evaluation of grammar Diabetes Register with Diabetes Type 2. The period of
coverage and extraction accuracy in general. Some of the interest is two years preceeding the onset of the Diabetes
main advantages of the proposed method, beyond its Type 2, i.e. the so called prediabetes condition. In order
reliable performance and good precision in text mining, to illustrate the potential of contextualized FPM we
are the modularity, extensibility, and scalability. present results in searching comorbidities for patients in
prediabet condition. Text mining modules are used to
convert raw text descriptions to structured event data.
233
𝐸(𝑝𝑖 ). Let 𝐷 ⊆ 𝑃 × 2𝐼 be the set of all itemsets in our
Outpatient collection after projection 𝜋 in the format
Records
〈𝑝𝑖𝑑, 𝑖𝑡𝑒𝑚𝑠𝑒𝑡〉. We shall call 𝐷 a database. We are
looking for itemsets 𝑋 ⊆ 𝐼 with frequency (sup(𝑋))
above given 𝑚𝑖𝑛𝑠𝑢𝑝. Let ℱ denote the set of all frequent
itemsets, i.e. ℱ = {𝑋| 𝑋 ⊆ 𝐼 𝑎𝑛𝑑 sup(𝑋) ≥ 𝑚𝑖𝑛𝑠𝑢𝑝}. A
Preprocessing
frequent itemset 𝑋 ∈ ℱ is called maximal if it has no
Structured
Text
Information
Data frequent supersets. Let ℳ denote the set of all maximal
Mining Modeling
Processing frequent itemsets, i.e. ℳ = {𝑋| 𝑋 ∈ ℱ 𝑎𝑛𝑑 ∄ 𝑌 ∈
ℱ, 𝑠𝑢𝑐ℎ 𝑡ℎ𝑎𝑡 𝑋 ⊂ 𝑌}. Let 2𝑋 denote the power set (set
of all subsets) of itemset 𝑋. Then each subset of 𝑋 ∈ ℱ is
Data Analysis
also a frequent itemset, i.e. ∀ 𝑌 ∈ 2𝑋 𝑖𝑚𝑝𝑙𝑖𝑒𝑠 𝑡ℎ𝑎𝑡 𝑌 ∈
ℱ . For each item 𝑖𝑑 ∈ 𝐼 we define the set called pidset:
Structured Association Context
Information Rules Information 𝑝(id) = {𝑝𝑖 | 〈𝑝𝑖 , 𝐼(𝑝𝑖 )〉 ∈ 𝐷 𝑎𝑛𝑑 𝑖𝑑 ∈ 𝐼(𝑝𝑖 )}.
Processing Generation Processing To study the nature of comorbidities we need to
investigate the context in which they occur. Therefore we
add some semantic attributes to each event –
Prediction & Prevention Models demographics of patients, age and gender, treatment,
status, lab data and etc.
Comorbidity Risk Factors We define a set of attributes of interest 𝐴 =
Analysis
{𝑎1 , 𝑎2 , … , 𝑎𝑘 }. Context Q for some patient 𝑝𝑖 ∈ 𝑃 is
defined as the set of attribute-value pairs from patient
Figure 5 System Architecture profile information:
𝑄(𝑝𝑖 ) = {〈𝑎1 , 𝑞1 〉, 〈𝑎2 , 𝑞2 〉, … , 〈𝑎𝑘 , 𝑞𝑘 〉}.
The search space is very large: the database is big, the In order to decrease the number of possible values of
number of diseases is also large. We propose a tabular attributes we apply some aggregation of data. For
method using a vertical database, depth-first traversal as instance age value is categorized according to the World
well as set intersection and diffsets [19]. Further Health Organization (WHO) standard age groups. Data
processing of the maximal frequent itemsets (MFI) is for body mass index (BMI) are also categorized
applied to remove diagnostic-related groups. In addition according to the WHO4 standard classification -
some context information is added to each MFI to underweight, normal weight, overweight, obesity.
investigate comorbidities. Furthermore association rules For some data concerning demographic information,
with lift are generated. The context information is like region ID we have large number of distinct values.
represented as attribute-value tuples for each patient; the For such data we add also some additional properties
post-processing identifies the importance of different concerning background information for the region – e.g.
attributes for each MFI. whether it is south, north, west, east, central, northwest
The architecture of the experimental workbench is etc., and mountain, river, sea, thermal spring, urban
shown on Figure 5. Our research [19] aims to develop region etc. For status and clinical test data we take the
further the ideas of the two contextual approaches for worst value for the period, according to the risk factors
data mining [22, 23]. deinition.
For the collection S of ORs we extract the set of all In primary interest for Diabetes Type 2 are BMI,
different patient identifiers 𝑃 = {𝑝1 , 𝑝2 , … , 𝑝𝑁 }. This set glycated haemoglobin, blood pressure (RR – Riva Roci),
corresponds to transaction identifiers (tids) and we call blood glucose, HDL-cholesterol.
them pids (patient identifiers). We consider each patient From 𝑄(𝑝𝑖 ) we generate a feature vector 𝑣(𝑝𝑖 ) =
visit to a doctor as a single event. For each patient 𝑝𝑖 ∈ 𝑃 (𝑣1𝑖 , 𝑣2𝑖 , … , 𝑣𝑚𝑖 ), where each attribute 𝑎𝑗 ∈ 𝐴 with 𝑁𝑗
an event sequence of tuples 〈𝑒𝑣𝑒𝑛𝑡, 𝑡𝑖𝑚𝑒𝑠𝑡𝑎𝑚𝑝〉 is possible values is represented by 𝑁𝑗 consecutive
generated: 𝐸(𝑝𝑖 ) = (〈𝑒1 , 𝑡1 〉, 〈𝑒2 , 𝑡2 〉, … , 〈𝑒𝑘𝑖 , 𝑡𝑘𝑖 〉), 𝑖 = positions in the vector. For the set of maximal frequent
̅̅̅̅̅
1, 𝑁. Let ℰ be the set of all possible events and 𝒯 be the itemset ℳ with cardinality |ℳ| = K we have K classes
set of all possible timestamps. Let 𝐼 = {𝑖𝑑1 , 𝑖𝑑2 , … , 𝑖𝑑𝑝 } of comorbidities. We apply classification of multiple
be the set of all diseases ICD-103 codes, which we call classes in order to generate rules for each comorbidity
items. Each subset 𝑋 ⊆ 𝐼 is called an itemset. We define class. We use large scale multi class classification
a projection function 𝜋: (ℰ × 𝒯)𝑁 → 2𝐼 : 𝜋(𝐸(𝑝𝑖 )) = because we deal with a big database and a large group of
𝐼(𝑝𝑖 ) = (𝑖𝑑1i , 𝑖𝑑2i , … , 𝑖𝑑𝑚𝑖 ), such that for each patient comorbidity classes. We use Support Vector Machines
(SVM) and optimization based on block minimization
𝑝𝑖 ∈ 𝑃 the projected time sequence contains only the
method described by Yu et al. [24].
first occurrence (onset) of each disorder recorded in
3 4
International Classification of Diseases and Related WHO, BMI Classification http://apps.who.int/bmi/
Health Problems 10th Revision. http://apps.who.int/ index.jsp?introPage=intro_3.html
classifications/icd10/browse/2015/en
234
Table 1. Data analysis results for patients in prediabetes condition
Set 2013 2014 2013-2014
ICD-10 ICD-10 ICD-10 ICD-10 ICD-10 ICD-10
items
3 signs 4 signs 3 signs 4 signs 3 signs 4 signs
Patients 27,082 27,082 27,902 27,902 29,205 29,205
Outpatient Records 267,194 267,194 296,129 296,129 556,323 556,323
ICD-10 codes 1,142 4,701 1,145 4,834 1,257 5,503
minsup 0.01 0.01 0.01 0.01 0.01 0.01
Total MFI 203 486 219 512 521 1,406
Longest MFI 5 8 5 9 6 9
Frequent Itemsets 608 7,452 689 8,935 1,909 32,093
Association Rules 686 58,299 810 78,052 2,722 381,012
5.3 Experiments and Results
Table 2. Data for attributes in the collections
We report results for patients with Diabetes Type 2
onset in 2015. The ORs of these patients for the period 𝐴 attribute 2013 2014 2013-2014
2013-2014 were excerpted from the Diabetes Register 𝑎1 age 27,082 27,902 29,205
when, as we assume, these patients were in a pre-diabetes 𝑎2 gender 27,082 27,902 29,205
condition. The idea of this experiment is to check 𝑎3 region 27,082 27,902 29,205
whether we can successfully discover risk factors for 𝑎4 bmi 21,659 22,413 27,928
these patients looking only at their ORs in 2013 and 𝑎5 HbA1c 153 238 370
2014. Then, maping our hypotheses to the real data for HDL
2015, we test whether our approach is reasonable. (We 𝑎6 4,917 4,815 6,952
cholesterol
note that due to the relatively short period of observation 𝑎7 blood glucose 11,925 12,185 17,016
and lack of data about mortality, at the moment we
cannot follow diabetes development in longer periods.)
In the Register each OR, corresponding to a single One of the generated Maximal Frequent Itemsets
visit, cointains up to 4 diagnoses encoded in ICD-10. (comorbidity class), whose support contains the pid=
Some diagnoses are presented by 4-sign encodings, i.e. 2196365, is:
in a more specific way, while others use the more general MFI#12: Z00 I10 M51 #SUP: 453
3-sign encoding. Due to the hierarchical organization of The following charts show the distribution of patients
ICD-10 we shall analyse individually two collections: in the support of "MFI#12" according to their age (Figure
the original one, that is more specific (with 4-sign codes 6), gender (Figure 7), BMI (Figure 8), and the HDL
- see Example 1) and we shall generalise also all Cholesterol (Figure 9) correspondingly.
diagnoses to more general classes (with 3-sign codes - We can observe that most patients in this support set
see Example 2). The examples present collections of have higher risk of Diabetes Type 2, due to the presence
diagnoses for a patient with ID 2196365. of multiple risk factors as obesity, medium or high levels
Example 1: of cholesterol and hypertension (diagnose with ICD-10
I(2196365)={I10, M10.9, M10, K76.9, K76, code I10).
L94.1, L94, M06.9, G57.9, Z00.8, H53,
M51.1, M33.9}
AGE
Example 2:
I(2196365)={I10, M10, K76, L94, M06, M51, 200
M33, H53, Z00, G57}
150
For some patients, the available ORs contain no
information about certain attributes of the context 100
information (Table 2). It is well known that missing data
in medical documentation is inevitable. Thus some 50
attribute values are replaced by the value NA, which is
considered as the most general value. 0
For example the context information for the patient 15-44 45-59 60-69 70-89
with ID 2196365 is: Figure 6 Age of the patients in the support set of
"MFI#12"
𝑄 (2196365) = {〈𝑎𝑔𝑒, 58〉, 〈𝑔𝑒𝑛𝑑𝑒𝑟, 1〉,
〈𝑟𝑒𝑔𝑖𝑜𝑛, 03〉, 〈𝑏𝑚𝑖, 29.32〉, 〈ℎ𝑏𝑎1𝑐, 𝑁𝐴〉,
〈𝑏𝑙𝑜𝑜𝑑_𝑔𝑙𝑢𝑐𝑜𝑠𝑒, 6.39〉, 〈ℎ𝑑𝑙_𝑐ℎ𝑜𝑙𝑒𝑠𝑡𝑒𝑟𝑜𝑙, 1.15〉}
235
with ICD-10 code M51 (Thoracic, thoracolumbar, and
GENDER lumbosacral intervertebral disc disorders) means that the
patients have lower motor activity and sedentary
lifestyle, which causes obesity, overweight, higher
male values of cholesterol and blood pressure and therefore
35% increases the risk of developing Diabetes. Actually this
has happened in 2015. We note that in general the ICD-
female 10 diagnose M51 is not considered risky for Diabetes.
65% But our algorithm reveals this unknown and latent
interrelationship.
6 Conclusion and Future Work
Figure 7 Gender of the patients in the support set of In this paper we present a software environment for
"MFI#12" collection and processing of Big Data in medicine - a
Data Intensive Domain. The Diabetes Register has been
developed stepwise and its research functionality is still
BMI underweight under construction. We believe that the integration of
0%
various technologies is the proper way to approach the
normal challenges of large-scale information processing because
9% the integration ensures flexible multi-functionality and
obesity enables reuse of results.
55% overweight The nation-wide Diabetes Register of Bulgaria is now
36% visible in Internet5 together with some public statistical
information. We plan to develop the Register further as a
predictive and preventing tool using the synergy of
advanced technologies which enable to discover risk
groups of patients that have predisposition to various
socially-significant diseases. We have shown here that
Figure 8 BMI of the patients in the support set of the present software environment is mature enough to
"MFI#12" identify patients with complexes of risk factors for
development of Diabetes, e.g. risks like: family history
(relatives with Diabetes); obesity; arterial hypertonia
HDL CHOLESTEROL (RR>140/90); low physical activity; giving birth to a
baby with weight more than 4 kg or gestational Diabetes;
established impaired fasting glycaemia or impaired
low high glucose tolerance; other states of insulin resistance (e.g.
21% 23% acanthosis nigricans, a specific hyperpigmentation of the
skin that might be due to endocrine disorders); HDL-
medium cholesterol≤0.90 mmol/l or triglycerides≥2.2 mmol/l
(≥2.82 mmol/l according to ADA); diagnosed polycystic
56%
ovarian syndrome, a cardio-vascular disease, or mental
disorders etc. These risk factors are explicated in the
patient-related documents either by values of clinical
tests or by keywords and typical phrases that describe the
Figure 9 Levels of HDL Cholesterol of the patients in factor. The patients with predisposition suffer from
the support set of "MFI#12" disorders and syndromes, diagnosed by various medical
specialists in various time periods, but without any
Data about HbA1c are available only for 3 out of 453 chance to establish connections between the medical
patients, that is why we consider this attribute as a more doctors – e.g. a connection between a Psychiatrist and a
general value ANY. But we note that the lack of HbA1c Cardiologist that have consulted the patient. Elaborating
measurements is not surprising because tests for HbA1c further the analytics facility of the Register will provide
are made when the Diabetes is diagnosed (and this has functionality to monitor patient status over time, in the
happened in 2015 for the selected patient cohort). context of all available information, and to issue alerts
Data for blood glucose are available only for 30% of for coincidence of risk factors that open the door to
these patient and for 50% of them the values were high. Diabetes and other chronic diseases. In this way we
Deeper analyses reveal medical arguments why believe that in the foreseeable future it will become
higher risk exist especially for the patients in the support possible to identify the Bulgarian citizens who have
set of MFI#12: Z00 I10 M51 #SUP: 453. The diagnose predisposition to develop Diabetes Mellitus.
5
http://usbale.com/Register_Diabetes.htm
236
Acknowledgements [13] Boytcheva, S.: Shallow Medication Extraction
from Hospital Patient Records. Studies in Health
The research presented here is partially supported by the Technology and Informatics. vol. 166, pp. 119-
grant 02/4 SpecialIZed Data MIning MethoDs Based on 128. IOS Press (2011)
Semantic Attributes (IZIDA), funded by the National
[14] Tcharaktchiev, D., Angelova, G., Boytcheva, S.,
Science Fund in 2017–2019. The support of Medical
Angelov, Z., Zacharieva, S.: Completion of
University – Sofia, the Bulgarian Ministry of Health and
Structured Patient Descriptions by Semantic
the National Health Insurance Fund is acknowledged.
Mining. Studies in Health Technology and
Informatics, vol. 166, pp. 260–269. IOS Press
References (2011). doi: 10.3233/978-1-60750-740-6-260
[1] Carstensen, B. et al.: The Danish National [15] Laney, D.: 3D Data Management: Controlling
Diabetes Register: Trends in incidence, prevalence Data Volume, Velocity, and Variety. META
and mortality. Diabetologia. 51(12), 2187–2196 Group Research Note, 6, 10 (2001)
(2008). doi: 10.1007/s00125-008-1156-z https://blogs.gartner.com/doug-laney/files/2012/
[2] Hallgren Elfgren, I. M., Grodzinsky, E., Törnvall, 01/ad949-3D-Data-Management-Controlling-
E.: The Swedish National Diabetes Register in Data-Volume-Velocity-and-Variety.pdf
clinical practice and evaluation in primary health [16] Top 238 Business Analytics Tools. Predictive
care. Prim. Health Care Res. Dev. 17(6), 549-558 Analytics Magazine (Feb 2012).
(2016). doi: 10.1017/S1463423616000098 http://www.predictiveanalyticstoday.com/top-
[3] Cooper, J. G., Thue, G., Claudi, T., Løvaas, K., business-intelligence-tools/
Carlsen, S., Sandberg, S.: The Norwegian Diabetes [17] Angelova, G., Nikolova, I., Angelov, Zh.:
Register for Adults – an overview of the first years. Embedding language technologies in a data
Norsk Epidemiologi. 23(1), 29-34 (2013) analytics tool. Advances in Bulgarian Sciences,
[4] O'Mullane, M., McHugh, S., Bradley, C. P.: pp. 29-42. National Centre for Information and
Informing the development of a national diabetes Documentation (2016). ISSN: 1314-3565
register in Ireland: a literature review of the impact [18] Nasreen, S., Azam, M. A., Shehzad, K., Naeem,
of patient registration on diabetes care. Inform. U., Ghazanfar, M. A.: Frequent Pattern Mining
Primary Care. 18(3), 157-68 (2010) Algorithms for Finding Associated Frequent
[5] Hallgren Elfgren, I.M., Törnvall, E., Grodzinsky, Patterns for Data Streams: A Survey. Procedia
E.: The process of implementation of the diabetes Computer Science, 37, 109-116 (2014)
register in Primary Health Care. Int. Journal of [19] Boytcheva, S., Angelova, G., Angelov, Z.,
Qual. Health Care. 24(4), 419-424 (Aug 2012) Tcharaktchiev, D.: Mining Comorbidity Patterns
[6] Meystre, S., Savova, G., Kipper-Schuler, K., Using Retrospective Analysis of Big Collection of
Hurdle, J. F.: Extracting Information from Textual Outpatient Records. Health Inf Sci Syst. Journal,
Documents in the Electronic Health Record: A Springer (2017). ISSN: 2047-2501 (to appear)
Review of Recent Research. IMIA Yearbook of [20] Tcharaktchiev, D., Zacharieva, S., Angelova, G.,
Medical Informatics, pp. 138-154. (2008) Boytcheva, S., et al. Building a Bulgarian National
[7] UMLS, the Unified Medical Language System. Registry of Patients with Diabetes Mellitus.
https://www.nlm.nih.gov/research/umls/ Journal of Social Medicine. 2, 19-21 (2015) (in
Bulgarian)
[8] Denny, J. C., Irani, P. R., Wehbe, F. H., Smithers,
J. D., Spickard, A.: The KnowledgeMap Project: [21] Boytcheva, S., Angelova, G., Angelov, Z.,
Development of a Concept-Based Medical School Tcharaktchiev, D.: Text Mining and Big Data
Curriculum Database. In: AMIA Annu Symp Analytics for Retrospective Analysis of Clinical
Proc., pp. 195–199. (2003) Texts from Outpatient Care. Cybernetics and
Information Technologies, 15(4), 58-77 (2015).
[9] Chapman, W., Bridewell, W., Hanbury, P.,
doi: 10.1515/cait-2015-0055
Cooper, G. F., Buchanan, B.: A Simple Algorithm
for Identifying Negated Findings and Diseases in [22] Rabatel, J., Bringay, S., Poncelet, P.: Mining
Discharge Summaries. Univ. of Pittsburgh (2002) sequential patterns: a context-aware approach.
Advances in Knowledge Discovery and
[10] Gindl, S.: Negation Detection in Automated
Management, pp. 23-41. Springer (2013)
Medical Applications. TUW (2006)
[23] Ziembiński, R. Z.: Accuracy of generalized
[11] HITEx Manual: https://www.i2b2.org/software/
context patterns in the context based sequential
projects/hitex/hitex_manual.html
patterns mining. Control and Cybernetics. 40, 585-
[12] Liao, K. P., Cai, T., Savova, G. K., Murphy, S. N., 603 (2011)
Karlson, E. W., Ananthakrishnan, A. N., Gainer,
[24] Yu, H. F., Hsieh, C. J., Chang, K. W., Lin, C. J.:
V. S. et al.: Development of phenotype algorithms
Large linear classification when data cannot fit in
using electronic medical records and incorporating
memory. ACM Transactions on Knowledge
natural language processing. British Med. J., 350
Discovery from Data (TKDD), 5(4), 23 (2012)
(1): h1885 (2015)
237