Content uploaded by Vadim Skobtsov
Author content
All content in this area was uploaded by Vadim Skobtsov on Jan 29, 2021
Content may be subject to copyright.
ISSN 1561-5359. Artificial Intelligence, 2020, № 3
© T.O. Bardadym, V.M. Gorbachuk, N.A. Novoselova, S.P. Osypenko, V.Yu. Skobtsov 65
UDC 519.8
T.O. Bardadym1, V.М. Gorbachuk1, N.А. Novoselova2, С.P. Osypenko1, V.Yu. Skobtsov2
1V.M.Glushkov Institute of Cybernetics, National Academy of Sciences of Ukraine, Ukraine
Academician Glushkov Ave., 40, Kyiv, 03187
2United Institute of Informatics Problems of the National Academy of Sciences of Belarus, Belarus
6, Surganava St., Minsk, 220012
INTELLIGENT ANALYTICAL SYSTEM AS A TOOL TO ENSURE
THE REPRODUCIBILITY OF BIOMEDICAL CALCULATIONS
Т.О. Бардадим1, В.М. Горбачук1, Н.А. Новоселова2, С.П. Осипенко1, В.Ю. Скобцов2
1Інститут кібернетики імені В.М. Глушкова НАН України, Україна
пр. Академіка Глушкова, 40, м. Київ, 03187
2Об’єднаний інститут проблем інформатики НАН Білорусі, Білорусь
вул. Сурганова, 6, м. Мінськ, 220012
ІНТЕЛЕКТУАЛЬНА АНАЛІТИЧНА СИСТЕМА ЯК ІНСТРУМЕНТ
ЗАБЕЗПЕЧЕННЯ ВІДТВОРЮВАНОСТІ
БІОМЕДИЧНИХ ОБЧИСЛЕНЬ
Abstract. The experience of the use of applied containerized biomedical software tools in cloud environment is
summarized. The reproducibility of scientific computing in relation with modern technologies of scientific calculations
is discussed. The main approaches to biomedical data preprocessing and integration in the framework of the intelligent
analytical system are described. At the conditions of pandemic, the success of health care system depends significantly
on the regular implementation of effective research tools and population monitoring. The earlier the risks of disease can
be identified, the more effective process of preventive measures or treatments can be. This publication is about the
creation of a prototype for such a tool within the project «Development of methods, algorithms and intelligent analy-
tical system for processing and analysis of heterogeneous clinical and biomedical data to improve the diagnosis of com-
plex diseases» (M/99-2019, M/37-2020 with support of the Ministry of Education and Science of Ukraine), implement-
ted by the V.M. Glushkov Institute of Cybernetics, National Academy of Sciences of Ukraine, together with the United
Institute of Informatics Problems, National Academy of Sciences of Belarus (F19UKRG-005 with support of the Bela-
russian Republican Foundation for Fundamental Research). The insurers, entering the market, can insure mostly low
risks by facilitating more frequent changes of insurers by consumers (policyholders) and mixing the overall health insu-
rance market. Socio-demographic variables can be risk adjusters. Since age and gender have a relatively small explana-
tory power, other socio-demographic variables were studied – marital status, retirement status, disability status, educati-
onal level, income level. Because insurers have an interest in beneficial diagnoses for their policyholders, they are also
interested in the ability to interpret relevant information – upcoding: insurers can encourage their policyholders to con-
sult with doctors more often to select as many diagnoses as possible. Many countries and health care systems use diag-
nostic information to determine the reimbursement to a service provider, revealing the necessary data. For processing
and analysis of these data, software implementations of construction for classifiers, allocation of informative features,
processing of heterogeneous medical and biological variables for carrying out scientific research in the field of clinical
medicine are developed. The experience of the use of applied containerized biomedical software tools in cloud environ-
ment is summarized. The reproducibility of scientific computing in relation with modern technologies of scientific cal-
culations is discussed. Particularly, attention is paid to containerization of biomedical applications (Docker, Singularity
containerization technology), this permits to get reproducibility of the conditions in which the calculations took place
(invariability of software including software and libraries), technologies of software pipelining of calculations, that
allows to organize flow calculations, and technologies for parameterization of software environment, that allows to re-
produce, if necessary, an identical computing environment. The main approaches to biomedical data preprocessing and
integration in the framework of the intelligent analytical system are described. The experience of using the developed
linear classifier, gained during its testing on artificial and real data, allows us to conclude about several advantages pro-
vided by the containerized form of the created application: it permits to provide access to real data located in cloud en-
vironment; it is possible to perform calculations to solve research problems on cloud resources both with the help of
developed tools and with the help of cloud services; such a form of research organization makes numerical experiments
reproducible, i.e. any other researcher can compare the results of their developments on specific data that have already
been studied by others, in order to verify the conclusions and technical feasibility of new results; there exists a univer-
sal opportunity to use the developed tools on technical devices of various classes from a personal computer to powerful
cluster.
Keywords: classifier; cloud service; containerized application; gene expression data; isolated software
environment; reproducibility of calculations; biomarker.
ISSN 1561-5359. Штучний інтелект, 2020, № 3
66 © Т.О. Бардадим, В.М. Горбачук, Н.А. Новоселова, С.П. Осипенко, В.Ю. Скобцов
Анотація. Підсумовано досвід використання прикладних контейнеризованих біомедичних програмних
засобів у хмарному середовищі. Вказано шляхи забезпечення відтворюваності наукових обчислень при вико-
ристанні сучасних технологій наукових розрахунків. Описано основні підходи до попередньої обробки та ін-
теграції біомедичних даних у рамках інтелектуальної аналітичної системи. В умовах пандемії успіхи системи
охорони здоров’я суттєво залежать від регулярного впровадження ефективних засобів досліджень і моніторин-
гу стану населення. Чим раніше вдається виявити ризики появи захворювання, тим ефективніше може йти про-
цес профілактичних заходів або лікування. У даній публікації йдеться про створення прототипу такого засобу
в рамках проєкту «Розробка методів, алгоритмів і інтелектуальної аналітичної системи для обробки й аналізу
різнорідних клінічних та біомедичних даних з метою вдосконалення діагностики складних захворювань»
(М/99-2019, M/37-2020 за підтримки Міністерства освіти та науки України), що виконується Інститутом кібер-
нетики імені В.М.Глушкова НАН України спільно з Об’єднаним інститутом проблем інформатики НАН Біло-
русі (Ф19УКРГ-005 за підтримки Білоруського республіканського фонду фундаментальних досліджень). Стра-
ховики, що входять у ринок, можуть страхувати переважно низькі ризики, сприяючи частішим змінам страхо-
виків з боку страхувальників і змішуючи загальний ринок страхування. Коригувачами ризику можуть бути
соціально-демографічні змінні. Оскільки вік і стать мають відносно невелику пояснювальну спроможність, то
вивчалися інші соціально-демографічні змінні − сімейний статус, пенсійний статус, статус інвалідності, освіт-
ній рівень, рівень доходу. Оскільки страховики мають інтерес до вигідних діагнозів для своїх страхувальників,
то також мають інтерес до можливостей трактування відповідної інформації − перекодування інформації: стра-
ховики можуть заохочувати своїх страхувальників консультуватися з лікарями, щоб відбирати більше діагно-
зів. Багато країн і систем охорони здоровʼя використовують діагностичну інформацію для визначення відшко-
дування провайдеру відповідних послуг, відкриваючи необхідні для цього дані. Для обробки й аналізу цих да-
них розробляються програмні реалізації побудови класифікаторів, виділення інформативних ознак, опрацю-
вання різнорідних медико-біологічних змінних для проведення наукових досліджень у галузі клінічної меди-
цини. У статті підсумовано досвід використання прикладних контейнеризованих біомедичних програмних за-
собів у хмарному середовищі. Вказано шляхи забезпечення відтворюваності наукових обчислень при викорис-
танні сучасних технологій наукових розрахунків. Зокрема, увага привертається до контейнеризації біомедич-
них додатків (технології Docker, Singularity), за рахунок чого досягається відтворюваність середовища для ви-
конання обчислень (використання ідентичних програмних засобів та бібліотек), технології конвеєризації, що
допомагає організувати обчислення в потоковому режимі, та технології параметризації обчислювального сере-
довища, що дозволяє, за необхідності, створювати ідентичне обчислювальне середовище. Описано основні під-
ходи до попередньої обробки та інтеграції біомедичних даних в рамках інтелектуальної аналітичної системи.
Досвід використання розробленого лінійного класифікатора, набутий при його тестуванні на штучних та ре-
альних даних, дозволяє зробити висновок про декілька переваг, які надає контейнеризована форма створеного
додатку: вдається забезпечити доступ до реальних даних, розташованих у хмарних середовищах; забезпечу-
ється можливість виконання обчислень для розв’язування дослідницьких задач на хмарних ресурсах як за до-
помогою розроблених засобів, так і за допомогою хмарних сервісів; така форма організації досліджень робить
числові експерименти відтворюваними, тобто будь-який інший дослідник може порівняти результати роботи
своїх розробок на конкретних даних, які вже вивчали інші, з метою перевірити зроблені висновки та технічні
можливості нових розробок; з’являється універсальна можливість використовувати розроблені засоби на тех-
нічних пристроях різного класу від персонального комп’ютера до потужного кластера.
Ключові слова: класифікатор; хмарний сервіс; контейнеризований додаток; дані експресії генів;
ізольоване програмне середовище; відтворюваність обчислень; біомаркер.
Introduction
This publication summarizes the expe-
rience of the use of applied containerized soft-
ware tools in cloud environment, which the
authors gained during the project «Develop-
ment of methods, algorithms and intellectual
analytical system for processing and analysis
of heterogeneous clinical and biomedical data
in order to improve the diagnosis of complex
diseases», accomplished by the team from the
United Institute of Informatics Problems of the
NAS of Belarus and V.M.Glushkov Institute
of Cybernetics of the NAS of Ukraine. The
main approaches and program tools for the
development of intellectual analytical system
are described.
Problem formulation
At the conditions of pandemic, the suc-
cess of health care system depends signifi-
cantly on the regular implementation of effec-
tive research tools and population monitoring
[1−2]. Decisions on which the peopleʼs lives
depend are regularly made not only by indivi-
duals, but also by legislative and executive in-
stitutions of power that implement the func-
tion of state health care system. These deci-
sions take into account the possibility of pre-
serving and prolonging human life by means
of scarce resources (for example, financial,
human, temporal resources). Such decisions
are made by government institutions to imple-
ment the functions of defence and security,
ISSN 1561-5359. Artificial Intelligence, 2020, № 3
© T.O. Bardadym, V.M. Gorbachuk, N.A. Novoselova, S.P. Osypenko, V.Yu. Skobtsov 67
law and order, macroeconomic management,
protection of property rights.
Analysis of modern researches and
publications
The countries, having a national health
service or national health insurance, usually
allow government agencies to make decisions
on orders for new products (pharmaceuticals,
therapies and medical devices). As a rule, in-
novations, that promote therapeutic treatment
with a lower probability of early death within
a certain risk group, predominate [3]. Because
such innovations involve additional costs,
cost-cutting innovations are often neglected.
For instance, providing a multimillion-
dollar mobile coronary unit can help treat pa-
tients with heart attacks quickly, significantly
reducing the number of lethal cases on the
way to a hospital. The long-term drug therapy
for patients with hypertension who use anti-
hypertensive drugs can also prevent heart at-
tacks, significantly supporting the economy of
research and development (R&D) in pharma-
ceuticals. The installation of dialysis equip-
ment for patients with chronic renal failure
promotes R&D in manufacturing medical
equipment.
Goal of the research
The earlier the risks of disease can be
identified, the more effective process of pre-
ventive measures or treatments can be [4].
Life-saving costs are borne not only in the
field of health care: in the field of transport, in
locations with a higher number of road acci-
dents, there are issues of improving the quality
of roads (not only road surface), which must
be met by local communities and government
agencies; in the field of transport, there are
also issues of proper arrangement of roads
within residential areas in order to reduce
speed of vehicles and to conduct permanent
video surveillance. Of course, the practical
realization of responses to those issues in-
volves certain expenditures of the local or
state budget.
Main results
In the field of environmental protection,
there are questions about ensuring the levels of
security systems for such dangerous enter-
prises as a nuclear power plant or a chemical
plant; if the level of security system is insuffi-
cient, an accident can occur threatening the
lives of millions of people. One of the conse-
quences of the 1986 Chornobyl disaster was
an increase in cancer cases, especially in
Ukraine and Belarus. In thermal power plants
burning coal, there are questions about the
cost of filters that can contain sulfur dioxide
and other harmful emissions into the atmo-
sphere. Such emissions increase the incidence
of respiratory diseases among the people.
In all the above issues, government ins-
titutions cannot make rational decisions with-
out a comprehensive and accurate assessment
of future gains (and losses) caused by the im-
plementation of a particular project, as well as
without comparison of such gains with the
present value of cost flow associated with the
project. It is important for decision makers to
measure gains and costs in the same units.
Since project costs are usually measured in
monetary terms, it makes sense to measure all
gains in monetary terms as well. Therefore,
the prolongation of life or improvement of hu-
man health, caused by the implementation of
project should also be measured in monetary
units. Since it is difficult to assess the status of
health and life for a human being in monetary
units, economists have developed alternative
methods for assessing the state of health and
human life.
Different approaches to economic health
assessment compare the benefits of medical
intervention with the costs of this intervention.
Gains from intervention can be measured by
physical units on a one-dimensional scale, mo-
netary units, units of cardinal utility function
reflecting the multidimensional concept of
health in a scalar index.
Since the 1990-s, several states in the
world have taken steps to increase competition
for their health care insurers, hoping to im-
prove efficiency in their fields of health insu-
rance and health care. Then the generalized
equality of price and marginal cost will mean
that competing health insurers will charge a
high premium for high risks and at the same
time a low premium for low risks: high risks
ISSN 1561-5359. Штучний інтелект, 2020, № 3
68 © Т.О. Бардадим, В.М. Горбачук, Н.А. Новоселова, С.П. Осипенко, В.Ю. Скобцов
are characterized by a relatively high expected
cost of treatment due to the high probability of
disease. As the state wants all its citizens to be
provided with health insurance, there are
issues of risk selection in health insurance
markets.
One way to ensure an universal access
to health insurance is to provide targeted sub-
sidies to the poorer strata of population to
cover insurance premiums. In practice, govern-
ments regulate premiums, effectively elimina-
ting the dependence of premium charged by
an insurer on risk: in the United States, for
example, premium regulation applies so called
a community rating. In addition, the German
and Swiss regulators typically require insurers
to follow an open enrollment policy and ac-
cept all the applications. In the United States
Medicare gives its beneficiaries a choice bet-
ween the Medicare Plan itself and competing
health care plans, which receive a capitation
payment for every policyholder.
Therefore, in the countries mentioned,
there is a natural incentive to risk selection. If
each person pays the same insurance pre-
mium, the insurer will expect losses with high-
risk individuals (of high-risk type) and gains
with low-risk individuals (of low-risk type).
The economic viability and balance of any
health insurer presumes a sufficient number of
low-risk persons insured: insurers try to attract
as many such persons as possible. Therefore,
under the pressure of competition, all the insu-
rers will take part in the collection of cream on
market (cream-skimming), attracting favo-
rable risks and avoiding adverse risks.
Risk selection can take many forms. On
the one hand, health insurers can implement
direct risk selection by influencing who would
sign the insurance contract: for example, the
insurers may not pay their attention to the draft
contract from a high-risk person. Individuals
who are likely to need some medical care may
be asked to sign a contract that provides addi-
tional discount services or outright payments.
On the other hand, indirect risk selection is the
development of payment packages or contrac-
ting with service provides that involve low-
risk individuals but do not involve high-risk
persons. Direct risk selection concerns the
problem of individual access to a service, and
indirect one − the quality problem.
The both forms of risk selection will oc-
cur only when insurers or their consumers
possess information about individual health
care costs. Direct risk selection require insu-
rers to be able to observe the characteristics of
physical persons that correlate with their ex-
pected costs − gender, age, social behaviour,
and so on. For instance, if healthy people use
the Internet more often, the risk selection stra-
tegy is to market insurance contracts online:
this way people do not have to know their type
of risk. However, people need to know their
type of risk in indirect risk selection: for
example, people need to know the likelihood
that they will use certain services. Such perso-
nal data allow insurers to develop payment
packages and attract service providers with
different types of risk.
Direct and indirect risk selection can
take place simultaneously: measures that ex-
clude one selection should not affect another.
For instance, if the benefit package is strictly
regulated, preventing indirect risk selection,
insurers may remain interested in attracting fa-
vorable risks and thus turn to another risk se-
lection − direct risk selection. On the contrary,
if insurers do not have the ability to select
risks directly, they retain the incentive to deve-
lop a benefit package that attracts low risks
and avoids high risks. Indirect risk selection is
closely related to the phenomenon of unfavo-
rable (adverse) selection in insurance markets,
which happens when policyholders have more
information about their type of risk in compa-
rison with their insurers. This phenomenon
takes place regardless of the actions of state.
At the same time, indirect risk selection is an
implication of state regulations for premiums.
To avoid unwanted behavior by insurers
in selecting risks, certain measures can be
taken based on the assumption of compulsory
health insurance, forcing them to cover high
risks by means of low risks.
First, open enrollment guarantees that
some insurers will take some high risks. At the
same time, legislation, regulation and repor-
ISSN 1561-5359. Artificial Intelligence, 2020, № 3
© T.O. Bardadym, V.M. Gorbachuk, N.A. Novoselova, S.P. Osypenko, V.Yu. Skobtsov 69
ting may prevent obvious opportunities for di-
rect risk selection: for example, the law may
limit the insurerʼs financial and other benefits
from taking low risks.
Second, the measure against indirect
risk selection is the regulation of benefit pack-
age. On the one hand, lower bounds of bene-
fits can be envisaged, forcing insurers to offer
benefits that are important for high risks (say,
for the treatment of different types of dia-
betes). On the other hand, upper bounds of
payments may prevent insurers from including
low-risk services (say, fitness center services)
in their contracts. In addition, certain types of
payments, that are convenient for risk selec-
tion, can be regulated by separate provisions.
However, the payment package includes sup-
ply of services from specific partners provided
for in the contract (say, subcontractors), which
may be selected by the insurer in question.
Such selection is especially important in Ma-
naged Care: for example, by involving many
sports medicine professionals, the insurer can
count on the attention of healthy lifestyle ad-
vocates (low-risk consumers).
Third, the measure of creating incen-
tives via additional payments to insurers, who
take high risks, and imposing financial sanc-
tions to insurers, who skim creams (favorable
risks), is a risk adjustment scheme (RAS). The
payments mentioned depend on such characte-
ristics observed as age and gender. The mea-
sure of reimbursing the share of actual costs
for medical treatment is a cost reimbursement
scheme (CRS). The idea of CRS is to reduce
gains from risk selection by decreasing the im-
pact of costs on the profits of insurers. At the
same time, the CRS reduces incentives of in-
surers to control their costs.
The RAS and CRS can be substantiated
by modeling risk selection. First of all, due to
various reasons insurers may differ in their
terms of insurance for population, the RAS
and CRS can create a competitive system
where the favorable risk structure of an insurer
does not give her a starting advantage. Be-
sides, the health insurance market may be de-
stabilized as new insurers enter the market and
move from high to low risks. The RAS and
CRS can reduce differences of insurers in pre-
miums, thereby reducing incentives to the
movement (transition).
The insurers, entering the market, can
insure mostly low risks by facilitating more
frequent changes of insurers by consumers
(policyholders) and mixing the overall health
insurance market. Because insurers, that have
entered the market earlier, would appear at
high risks, they eventually have to increase
their premiums or file for bankruptcy. In such
circumstances, insurers will have no incentive
to invest in proving effective payments.
Indeed, there is evidence of higher low-
risk mobility in the German health insurance
market, based on a comparison of the health
care expenditure (HCE) of those who change
insurers and those who do not change their
insurer: depending on age categories, people,
who changed insurers, had on average
45−85 % less HCE than the HCE of those
who did not change insurers. Studies, based on
the German socioeconomic panel, have shown
that (adult) people, who remained loyal to
their insurer, had significantly worse health
status than people who changed insurers. In
the United States, there is a case of Harvard
University’s decision to increase employers’
contributions to insurance premiums if emplo-
yers did not choose the cheapest option
(Health Maintenance Organization (HMO)
plan). Types of risk began to be identified du-
ring the year: those who switched from the
most expensive insurance plans to HMOs had
a mean age of 46 years and were 9% higher in
HCE than the overall average HCE; those who
remained on expensive insurance plans had an
average age of 50 years and a 16% higher
HCE compared to the general average HCE.
The rapid loss of low risks by broad insurance
plans forced the experiment to stop.
Thus, the RAS and CRS can help ensure
a level playing field during the transition to a
competitive market and the stabilization of
health insurance market. In the absence of
schemes such as the RAS and CRS, the mar-
ket may lose the most efficient insurers. For
actuaries and other financial professionals, risk
ISSN 1561-5359. Штучний інтелект, 2020, № 3
70 © Т.О. Бардадим, В.М. Горбачук, Н.А. Новоселова, С.П. Осипенко, В.Ю. Скобцов
adjustment means the accrual of a premium or
per capita payment in proportion to the ex-
pected expenses of an individual or group. The
RAS is based upon risk adjusters – the ob-
served characteristics of individuals. The de-
velopment of RAS and the search for appro-
priate risk adjusters require empirical testing
of their ability to predict HCE.
Socio-demographic variables can be risk
adjusters. Since age and gender have a relati-
vely small explanatory power, other socio-de-
mographic variables were studied – marital
status, retirement status, disability status, edu-
cational level, income level. Data from the
German health insurance funds showed that
elderly pensioners with disabilities have signi-
ficantly higher HCE. In addition, higher HCEs
are revealed by single retirees and low-income
individuals.
HCE in previous periods is an obvious
indicator of morbidity: an increase in HCE
leads to an increase in HCE in the next period
by 20−30%. At the same time, the explanatory
capacity of HCE should be weighed against
the weakening of person’s incentives to reduce
her costs, because higher current HCE will to
some extent be compensated to the person la-
ter. It is through HCE that insurers try to iden-
tify favorable risks, and there may not be bet-
ter risk adjusters. Prescription medications in
previous periods have predicted the value of
HCE. The morbidity can be measured by ga-
thering available diagnostic information to
identify chronically ill patients and to classify
individuals according to their expected HCE.
This classification can be done by various me-
thods. The empirical studies show that diag-
nostic information gives an accurate predict-
tion of HCE values. In turn, the corresponding
gathering of information can be expensive.
Because insurers have an interest in beneficial
diagnoses for their policyholders, they are also
interested in the ability to interpret relevant in-
formation – upcoding: insurers can encourage
their policyholders to consult with doctors
more often to select as many diagnoses as
possible.
Many countries and health care systems
use diagnostic information to determine the re-
imbursement to a service provider, revealing
the necessary data. For processing and analy-
sis of these data, software implementations of
construction for classifiers, allocation of infor-
mative features, processing of heterogeneous
medical and biological variables for carrying
out scientific research in the field of clinical
medicine are developed.
One of the goals of research includes the
development of approaches and program tools
for the purpose of the reproducibility of nume-
rical experiments, which were conducted in
the framework of the joint project. The goal of
the project is to develop effective methods and
software for constructing classifiers, selection
of informative features, creation of a prototype
of an intelligent analytical system, which is a
software implementation of all stages of data
processing and analysis and is aimed at con-
ducting research in the field of clinical medi-
cine. This system will implement the functions
of integrating clinical and molecular patient
data, determining diagnostic biomarkers and
their combinations, building classifiers of
complex diseases (oncological diseases) based
on integrated data, identifying new disease
subtypes to improve treatment methods and
increase its efficiency. The second goal in-
cludes the development of the approaches to
Large amount of research activities de-
voted to the development of mathematical me-
thods of data handling, particularly classifica-
tion models, is due, on the one hand, to a wide
range of possible applications, and on the
other hand – the complexity of these prob-
lems, which requires the development and im-
provement of means to solve them (see, for
example, [5−9]). In addition to general requi-
rements for efficiency of the created software
there exists a need to pay attention to the con-
ditions of availability of large and heteroge-
neous data sets, requirements for the ability to
transfer programs from one hardware to
another, their performance in cloud
computing.
ISSN 1561-5359. Artificial Intelligence, 2020, № 3
© T.O. Bardadym, V.M. Gorbachuk, N.A. Novoselova, S.P. Osypenko, V.Yu. Skobtsov 71
Moreover, one of the most important re-
quirements is the reproducibility of research
numerical experiments. The principle of repro-
ducibility of research is one of the basic scien-
tific principles. However, a crisis called "re-
producibility crisis" has been realized in sci-
ence [10−11]. This crisis has affected almost
all branches of science, in particular, to a large
extent – biology and medicine. Much effort
has been made recently to overcome this cri-
sis, including the development of software and
software platforms to ensure the reproducibi-
lity of scientific computing. Computing in bio-
logy and medicine involves the use of high-
performance computing technologies (inclu-
ding clusters and grid technologies). However,
the introduction of modern technologies to en-
sure the reproducibility of calculations in this
area is quite slow [12, p. 731]. As a result, in
the field of cluster technologies, which do not
have the appropriate software installed, there
is a contradiction between modern require-
ments for the reproducibility of scientific cal-
culations and the ability to achieve it by old
means.
It so happened that the need to create a
containerized application was not a planned
stage of our study. This was primarily due to
the ways of accessing the real data on which
the software was tested. Only then did the
authors realize that they had gained other ad-
vantages, among which the most important is
the reproducibility of research numerical ex-
periments. It is the purpose of the publication
to share this experience.
The second purpose is to shortly de-
scribe our efforts taken towards the develop-
ment of specialized computer methods and
models in order to solve the vital tasks in the
field of biomedicine. Nowadays there exists
the enormous amount of biomedical and clini-
cal data collected in the public and private re-
positories. They can be freely accessed and
present the wide field for experiments with the
newly developed scientific approaches and
their comparison. The integration of heteroge-
neous information sources is one of the urgent
applied problems, which we have tried to
solve in our project. The hybrid classification
model presents the basis of the intelligent ana-
lytical system and aims to integrate several
sources of biomedical information in order to
improve the diagnostics and prognosis of
complex diseases.
Based on the approaches presented in
[13−14], optimization models and methods for
solving problems of constructing linear classi-
fiers have been developed. In particular, the
problem of constructing classifiers for linearly
indivisible sets was formulated as a problem
of minimizing the band of incorrect classifica-
tion of training sample points. This model be-
longs to the class of optimization problems of
non-convex programming and is multi-ext-
reme. Various formulations of this problem
are offered, approaches to construction of ap-
proximate decisions and calculation of estima-
tions of optimum values are considered. An
interesting geometric interpretation of the
problems of constructing linear classifiers can
be found in [15].
To solve these optimization problems,
methods of non-smooth optimization, namely
r-algorithms of N.Z. Shor [16−17] and exact
penalty functions [18–19] were used. When
creating appropriate software, modern libra-
ries of linear algebra, similar to [20−22]
should be used to speed up arithmetic opera-
tions. It is a combination of algorithms based
on non-smooth optimization methods and the
use of modern libraries of linear algebra was
implemented in the developed software mo-
dule NonSmoothSVC.
To test the abilities of the new classifier
NonSmoothSVC a comparison with existing
tools was made. The methods integrated into
the library scikit-learn [12; 23] were chosen,
namely Linear SVC, NuSVC, Ada Boost. The
two last methods are non-linear classifiers;
they were chosen to get additional information
concerning advantages of different methods
for different problems. First numerical experi-
ments were made on specially generated artifi-
cial data.
Computational experiments aimed to es-
tablish the speed and predictive properties of
new software compared to existing ones. Both
artificially created data and real medical data
were used in the calculations in the test prob-
lems. Training and control samples of ran-
ISSN 1561-5359. Штучний інтелект, 2020, № 3
72 © Т.О. Бардадим, В.М. Горбачук, Н.А. Новоселова, С.П. Осипенко, В.Ю. Скобцов
domly generated problems were formed as
identically distributed data points on a single
cube in the space of features
n
R
. Then, the
points of the first class shifted in the first coor-
dinate by the value δ, and the points of the se-
cond class shifted in the first coordinate by the
value (-1-δ). When δ>0, training and control
samples are linearly separable, and when δ<0,
they are linearly inseparable. Next, the rotation
(linear transformation) of space was per-
formed so that the separating hyperplane de-
pended on many coordinates of space. The
need to test new software on real data forced
us to locate the software module
NonSmoothSVC into a containerized applica-
tion (using Docker technology [24]) for use on
a personal computer, as well as on a cluster,
grid, and cloud environment.
Fig. 1. Comparative density distribution
for full data set (n=12000)
This permitted to get access to the real
data on Cancer Genomics Cloud [25], a speci-
alized cloud platform that provides free access
to genetic, medical databases, in particular –
The Cancer Genome Atlas (TCGA) [26], and
more than 450 public applications designed to
analyze data on this topic. It is possible to ex-
pand this list with the own applications, data
sets, research results (currently there are more
than one million on this service), to involve
other researchers in projects. Computational
experiments have demonstrated that on some
data sets the NonSmoothSVC has qualitative
advantages over other methods involved in the
comparison, but is inferior in speed. Parti-
cularly, on linearly separable samples the
NonSmoothSVC gained an advantage over the
LinearSVC in the number of cases with better
classification accuracy.
Fig. 2. Comparative density distribution
for imbalanced data set (n=3720)
On the unbalanced samples, the
NonSmoothSVC software slightly outperfor-
med the LinearSVC software in the number of
cases with better classification accuracy on
average, but demonstrated an advantage in
some parts of the classification accuracy scale
(Fig. 1–6).
Full description of numerical experi-
ments and the results of testing can be found
in the reports (in Ukrainian) at
http://moderninform.icybcluster.org.ua/ais/.
ISSN 1561-5359. Artificial Intelligence, 2020, № 3
© T.O. Bardadym, V.M. Gorbachuk, N.A. Novoselova, S.P. Osypenko, V.Yu. Skobtsov 73
Thanks to the containerized form, the
developed software can become publicly avai-
lable tools and applications of this and other
services in the problems of constructing opti-
mized linear classifiers using modern libraries
of linear algebra.
Fig. 3. Comparative density distribution
for large data set (n=7200)
In the presence of technical possibilities,
parallelization on microprocessor networks
looks promising.
This approach is especially recom-
mended in the case of large data samples, when
the dimension of the feature space is tens of
thousands. It is also necessary to take into
account the features of optimization problems
in specific cases. In particular, additional requi-
rements that may be formulated by specialists
may reduce the number of informative features.
Processing and study of biomedical data
have some peculiarities. This, in particular, the
existence of possible large errors that arise in
the processing of medical information and
huge number of features that need to be taken
into account, which increases the dimensiona-
lity of the corresponding optimization prob-
lems, the missed measurements, which requi-
res the use of specialized methods for their
processing and analysis.
In order to improve the diagnosis and
treatment of complex diseases, much attention
is paid to the comprehensive analysis of vari-
ous biomedical and clinical data to understand
the processes occurring in the body at the cel-
lular level and changes caused by the develop-
ment of the disease.
It is known, the cause of complex disea-
ses, along with external factors, is a combina-
tion of genetic failures, which does not allow
to fix only one genetic mutation as a biomar-
ker. The difficulty also lies in the fact that in-
dividual genetic factors can differ and indivi-
dual cases of the same disease (phenotype)
can be caused by different genetic changes. In
addition, in the case of the combined effect of
several mutations, the individual effect of each
of them can be rather insignificant and, there-
fore, difficult to be detected.
It is also necessary to take into account
the high heterogeneity of the complex disease,
i.e. heterogeneity of its observed manifesta-
tions (phenotypes).
Recently, the methods of systems bio-
logy have become widely used to study comp-
lex diseases, namely, knowledge about the
interactions between genes, their products and
small molecules that form a complex network
of interactions. This approach makes it pos-
sible to explain the appearance of similar phe-
notypes despite different genetic causes, na-
mely, their interconnection and influence (dys-
regulation) on the same component of the cel-
lular system. Thus, the use of interactome in
conjunction with other data from biogenetic
studies can contribute to understanding the
processes occurring at the molecular level in
complex diseases. The use of combinations of
heterogeneous data makes it possible to deter-
mine dysregulated cellular pathways, to reveal
the relationship between genotype and pheno-
type, and to explain the heterogeneity of a
complex disease.
ISSN 1561-5359. Штучний інтелект, 2020, № 3
74 © Т.О. Бардадим, В.М. Горбачук, Н.А. Новоселова, С.П. Осипенко, В.Ю. Скобцов
Fig. 4. Comparative density distribution
for small data set (n=2400)
Natural approaches here are: to increase
the efficiency of tools to solve such optimiza-
tion problems and the use of methods for se-
lection informative features. In the works
[27−30] attention is paid to the preliminary
preparation of available medical data in order
to select informative features.
In the course of the project, algorithms
for preprocessing and extracting biomarkers
from biomedical data were developed, inclu-
ding: an algorithm for ranking features by in-
formation content for classification [23]; an
algorithm for identifying combinations of bio-
markers, taking into account the correlation of
features and allowing to exclude their
influence.
Fig. 5. Comparative density distribution
for balanced data set (n=8280)
Fig. 6. Comparative density distribution
for imbalanced+small data set (n=720)
ISSN 1561-5359. Artificial Intelligence, 2020, № 3
© T.O. Bardadym, V.M. Gorbachuk, N.A. Novoselova, S.P. Osypenko, V.Yu. Skobtsov 75
Moreover, several approaches were ana-
lyzed for identifying a subset of informative
features, taking into account several data sour-
ces, namely, gene expression data and data on
functional and physical interactions of genes
and their products, presented in the form of
networks. Based on the analysis of existing
approaches, an algorithm for identifying a
subset of features has been developed, which
allows integrating interactomic and transcrip-
tomic data to determine functional subnets as-
sociated with the disease. Pre-processing of
biomedical data made it possible to reduce the
feature space and thereby increase the accu-
racy of classification models.
Detailed description of algorithms and
related information can be found in the report
at http://moderninform.icybcluster.org.ua/ais/
(in Russian).
In one of the numerical experiments the
real data contained information on the gene
expression of cancer patients (143 observa-
tions of 60,483 features) obtained from the
Cancer Genome Atlas (TCGA). From these
data by means of the simplified method of ran-
king of features proposed by Novoselova [28]
23 most informative features concerning the
forecast of a vital status of patients having
diagnosed glioblastoma were identified. This
approach substantially simplifies numerical
difficulties in following data processing.
Due to the fact that various sources of
biological information characterize various
changes occurring in the body at the cellular
level during the development of a complex
disease, it is assumed that their combination
will improve the accuracy of diagnosis of the
subtype of the disease, the reliability of the
disease prognosis and response to therapy
[31]. In addition, combining heterogeneous
data will allow one to discover the relation-
ships between various biomedical entities (ge-
nes, proteins, metabolites, etc.) directly related
to the development of the disease, compensate
for noise and errors in individual data sources
and thereby obtain more reliable results.
A common problem in solving this problem is
how to combine information from different
data sources. In our study, of interest are me-
thods for constructing classifiers based on va-
rious sources of multidimensional data, which,
as a rule, have a heterogeneous representation.
Consequently, the task is to unify this repre-
sentation, determine the base classifier, build
classification models on each data source, and
select ways to combine the predicted values,
obtained using the constructed models.
The core of the intelligent analytical
system being developed is a hybrid classifica-
tion model, which allows combining several
sources of biological information about pati-
ents in order to build a classification model
that allows diagnosing subtypes of complex
diseases characterized by genetic disorders.
The proposed hybrid model is a classification
ensemble with the following distinctive
features:
1. Uniform presentation of information from
various data sources by constructing a mat-
rix of object-object distances using various
kernel functions (density functions), inclu-
ding Gaussian, polynomial function, scalar
product of vectors, etc.
2. Implementation of the procedure for selec-
ting classification characteristics for each
individual data source.
3. Construction of a basic or individual classi-
fier of a hybrid model, which can be either
a single classifier or an ensemble of classi-
fiers built on a single data source.
4. Implementation of several ways of integra-
ting individual classifiers of the model.
5. Analysis of the information content of indi-
vidual classifiers using the assessment of
their weight coefficients.
The method for constructing a hybrid
model is based on a combination of the bag-
ging procedure and the aggregation of ranked
lists to build basic classifiers and a pruning
procedure to determine the final structure of
the model, which allows adaptively adjusting
the ensemble taking into account the type of
classified data.
The preliminary experiments on the
TCGA data [26] showed that the ensembles
built on heterogeneous data sources can suffi-
ISSN 1561-5359. Штучний інтелект, 2020, № 3
76 © Т.О. Бардадим, В.М. Горбачук, Н.А. Новоселова, С.П. Осипенко, В.Ю. Скобцов
ciently increase the accuracy of classification
and prediction of subtypes of complex disea-
ses, since each of the data sources describes
the organism under study in different planes:
gene expression data, Ribonucleic acid (RNA)
sequencing, metabolic data, gene copy number
data, etc.
Ensuring the reproducibility of calcula-
tions is a prerequisite for the reproducibility of
scientific research as a whole. The conditions
for computational reproducibility are the avai-
lability of source data, the ability to reproduce
an identical computing environment (or an en-
vironment that does not lead to other calcula-
tion results), and the availability of the results
of computations. Biomedical calculations have
their own specific features that should be
taken into account when planning them. Let
we mention some of them.
Modern biomedical calculations, espe-
cially based on genome data, are very huge
and cumbersome. Usually "classic" biome-
dical applications (PAML, Muscle, MAFFT,
MrBayes, BLAST, etc.) and large libraries
with implementations of biomedical algo-
rithms written in different programming lan-
guages (C / C ++, Java, R, Go, Scala, Haskell,
Perl, Python, Ruby, Erlang, Julia, etc. [32]) are
quite often used simultaneously in one study.
Moreover, biomedical calculations often in-
volve methods of artificial intelligence – ma-
chine learning, pattern recognition, and corres-
ponding libraries (e.g., scikit-learn [6], [17]).
Such a variety of software requires careful
configuration of the computing environment
with control of the versions of libraries used
(here can be used as dozens and hundreds of
libraries).
Otherwise one can get a lack of reprodu-
cibility as a result of calculations. In terms of
using cluster technologies, creating such envi-
ronments (separate for each user) and maintai-
ning them in a conflict-free state is quite a bur-
densome task (unless you use special software
configuration tools, such as Conda, Bioconda,
or containerization of applications using, for
example, technology Singularity). Most of the
libraries and applications used in biomedical
computing do not provide efficient use of pa-
rallel multithreaded computing with multi-
core processors, and at the same time many of
them can be applied to an "embarrassingly pa-
rallel" model − a model in which individual
pieces of data are calculated in parallel by
identical instances of computational processes
without transferring messages between them
(for example, using Apache Hadoop
technology) [12].
Taking into account the peculiarities of
biomedical computing, reproducibility and
their horizontal scaling (the ability to increase
the number of identical computing units to
solve one problem) can be achieved through
the use of containerized applications, software
pipeline computing and parameterization of
software environment.
Technologies of containerization of soft-
ware applications. Due to the containerization
of biomedical applications (Docker, Singula-
rity containerization technology) the following
can be achieved: reproducibility of the con-
ditions in which the calculations took place
(invariability of software including software
and libraries), the possibility of horizontal sca-
ling provided the use of "stunning" model of
parallelism in cluster (Singularity) and cloud
(using Docker) calculations.
Technologies of software pipelining of
calculations. Software pipeline allows you to
organize flow calculations (calculations in
which the inputs and outputs of processes are
interconnected). Thanks to the use of tools for
automation of flow calculations (workflow en-
gine) such as CWL (Common Workflow Lan-
guage), GWL (Guix Workflow Language),
Snakemake, Nextflow, it is possible to present
a specific calculation in the form of a task
(text file, as usual, in YAML format or
JSON), the results of which can be reproduced
[7]. In addition, there are tools that allow you
to create / display such tasks in the form of a
graph of processes and data flows. An
example of such a tool is RABIX (Reprodu-
cible Analyzes for Bioinformatics) − a graphi-
cal editor for CWL. Some pipeline tools also
use containerization (for example, CWL) −
ISSN 1561-5359. Artificial Intelligence, 2020, № 3
© T.O. Bardadym, V.M. Gorbachuk, N.A. Novoselova, S.P. Osypenko, V.Yu. Skobtsov 77
such tasks can be performed both on a perso-
nal computer and in a cloud environment. An
important feature of streaming automation
tools is that the task description syntax allows
you to specify the scale of the calculations, in-
dicating the number of resources required.
Seven Bridges' product, Cancer Genomics
Cloud (CGC, see
http://www.cancergenomicscloud.org/), is an
example of a cloud software platform for per-
forming reproducible biomedical computa-
tions using containerization and pipelining. It
is the use of containerization in the creation of
an application for the construction of a linear
classifier at the V.M. Glushkov Institute of
Cybernetics of the National Academy of Sci-
ences of Ukraine made it possible to conduct
testing on real very voluminous medical data
located at the CGC.
Technologies for parameterization of
software environment. Parameterization of the
software environment allows you to repro-
duce, if necessary, an identical computing en-
vironment. GNU Guix, Conda, Bioconda are
examples of tools that allow you to create an
isolated software environment for individual
users in a cluster [12].
At present, there exists a range of tech-
nologies to ensure the reproducibility of scien-
tific calculations in cloud and cluster environ-
ments. This makes it possible to create biome-
dical applications adapted to these environ-
ments. In the result we get computational basis
that satisfies modern requirements for compu-
tational reproducibility.
The experience of using the developed
linear classifier, gained during its testing on
artificial and real data, allows us to conclude
about several advantages provided by the con-
tainerized form of the created application: it
permits to provide access to real data located
in cloud environment; it is possible to perform
calculations to solve research problems on
cloud resources both with the help of deve-
loped tools and with the help of cloud ser-
vices; such a form of research organization
makes numerical experiments reproducible,
i.e. any other researcher can compare the re-
sults of their developments on specific data
that have already been studied by others, in or-
der to verify the conclusions and technical fea-
sibility of new results; there exists an universal
opportunity to use the developed tools on
technical devices of various classes from a
personal computer to powerful cluster.
Conclusions
The next steps of the project include
development of the common software inter-
face of the experimental prototype of the intel-
ligent analytical system in order to integrate
the developed methods and software modules
of biomedical data preprocessing, data cluste-
ring and classification. It will allow perfor-
ming all the steps of data analysis from the
single framework and conducting research in
the field of biomedicine. The hybrid classifica-
tion model as a core of the intelligent system
will make it possible to integrate multidimen-
sional, heterogeneous biomedical data with the
aim to better understand the molecular courses
of disease origin and development, to improve
the identification of disease subtypes and dise-
ase prognosis. Much attention will be paid to
the experimentation with different computa-
tion approaches on real datasets taking into ac-
count the reproducibility of results.
References
1. Knopov P.S., Norkin V.I., Atoyev K.L.,
Gorbachuk V.M., Kyryliuk V.S., Bila H.D.,
Samosyonok O.S., Bogdanov O.V. (2020). Some
approaches to the use of stochastic models of
epidemiology to the COVID-19 problem. Kyiv:
V.M.Glushkov Institute of Cybernetics, Retrieved
from http://incyb.kiev.ua/archives/3988/dejaki-
pidhodi-vikoristannja-stohastichnih-modelej-
epidemiologii-do-problemi-covid-19/
(In Ukrainian).
2. Gorbachuk V., Gavrilenko S. (2020). Analysis for
dynamics of COVID-19 spreading in Ukraine and
neighboring countries on May 1–10, 2020. Global
and regional problems of informatization in society
and nature using 2020. Kyiv: National University
of Life and Environmental Sciences of Ukraine,
56–60. (In Ukrainian).
3. Gorbachuk V.M., Dunaievskyi M.S., Suleimanov
S.-B. (2020). Management and administration in
the field of health care services. Management and
administration in the field of services: selected
examples. T.Pokusa, T.Nestorenko (eds.) Opole:
Academy of Management and Administration,
268−279. (In Ukrainian).
ISSN 1561-5359. Штучний інтелект, 2020, № 3
78 © Т.О. Бардадим, В.М. Горбачук, Н.А. Новоселова, С.П. Осипенко, В.Ю. Скобцов
4. Gorbachuk V.M., Suleimanov S.-B., Batih L.O.
(2020). Decision making criteria in the branch of
health care. Measurement and control in complex
systems. Vinnytsia: VNTU, 149–151.
(In Ukrainian).
5. Vorontsov K.V. Mathematical methods of learning
by precedents (Machine Learning Theory) (in
Russian), Retrieved from:
http://www.machinelearning.ru/wiki/images/6/6d/
Voron-ML-1.pdf
6. Gupal A.M., Sergienko I.V. Symmetry in DNA.
Methods for Discrete Sequences Recognition.
Kyiv. Naukova Dumka (in Russian).
7. Baldi P., Hatfield W.G. (2011). DNA Microarrays
and Gene Expression. From Experiments to Data
Analysis and Modeling. Cambridge University
Press.
8. Kuhn M., Johnson K. (2013). Applied predictive
modeling. New York: Springer.
9. Heath L.S., Ramakrishnan N. (2010). Problem
solving handbook in computational biology and
bioinformatics. NY: Springer Science & Business
Media.
10. Ioannidis J. (2005). Why Most Published Research
Findings Are False. PLoS Medicine, vol. 2, no. 8,
p. 124.
11. Baker M. (2016). Reproducibility crisis? Nature,
vol. 26, no. 533, 353-66.
12. Strozzi F. et al. (2019). Scalable workflows and
reproducible data analysis for genomics.
Evolutionary Genomics, 2nd ed., New York, NY:
Humana Press, 723-745.
13. Zhuravlev Y., Laptin Y., Vinogradov A.,
Zhurbenko N., Lykhovyd O., Berezovskyi O.
(2017). Linear classifiers and selection of
informative features. Pattern Recogn. and Image
Anal., vol. 27, no. 3, 426-432.
14. Zhuravlev Y., Laptin Y., Vinogradov A. (2014).
Comparison of Some Approaches to Classification
Problems, and Possibilities to Construct Optimal
Solutions Efficiently. Pattern Recogn. and Image
Anal., vol. 24, no. 2, 189-195.
15. Zhurbenko N.G. (2020). Linear classifier and
projection on polytop. Cybern. Syst. Anal., vol. 56,
no. 3, 1-8.
16. Shor N.Z., Zhurbenko N.G. (1971). A minimization
method using the operation of extension of the
space in the direction of the difference of two
successive gradients. Cybernetics, vol. 7, 450-459.
17. Shor N.Z. (1998). Nondifferentiable Optimization
and Polynomial Problems. London: Kluwer Acad.
Publ.
18. Laptin Yu.P. (2016). Exact penalty functions and
convex extensions of functions in decomposition
schemes in variables. Cybernetics and Systems
Analysis, vol. 52, 85–95. DOI: 10.1007/s10559-
016-9803-8.
19. Laptin Yu.P., Bardadym T.A. (2019). Problems
related to estimating the coefficients of exact
penalty functions. Cybernetics and Systems
Analysis, vol. 55, no. 3, 400-412. DOI:
10.1007/s10559-019-00147-2.
20. Chang, Chih-Chung; Lin, Chih-Jen LIBSVM –
A Library for Support Vector Machines. Retrieved
from https://www.csie.ntu.edu.tw/~cjlin/libsvm/.
21. BLAS (Basic Linear Algebra Subprograms).
Retrieved from http://www.netlib.org/blas/.
22. LAPACK—Linear Algebra PACKage. Retrieved
from http://www.netlib.org/lapack/.
23. Free software machine learning library for the
Python programming language. Retrieved from
https://scikit-learn.org/stable/index.html
24. Tools for creation of isolated Linux-containers.
Retrieved from https://www.docker.com/
25. The Cancer Genomics Cloud. Retrieved from
http://www.cancergenomicscloud.org/
26. The Cancer Genome Atlas (TCGA). Retrieved
from https://www.cancer.gov/about-
nci/organization/ccg/research/structural-
genomics/tcga
27. Novoselova N.A., Tom I.E. (2018). Integrated
network approach to protein function prediction.
The Scientific Journal of Riga Technical
University. Information Technology and
Management Science, vol. 21, 98–103. DOI:
10.7250/itms-2018-0016
28. Tom I.E. (2016). Information technologies in the
analysis of medical data. Science and innovations,
no. 3, 28-31.
29. Novoselova N.A., Tom I.E. (2016). Method for
constructing clusters in genetic data. Informatika,
no.1(49), 64-74.
30. Novoselova N.A., Tom I.E. (2013). Algorithm for
ranking features for detecting biomarkers in gene
expression data. Artificial Intelligence, no. 3, 58-
68.
31. Novoselova N.A., Tom I.E., Ablameyko S.V.
(2011). Evolutionary design of the classifier
ensemble. Artificial Intelligence, no. 3, 429-438.
32. Bonnal R. et al. (2019). Sharing Programming
Resources Between Bio* Projects. Evolutionary
Genomics, 2nd ed., New York, NY: Humana
Press, 747-766. DOI: 10.1007/978-1-4939-9074-
0_25
Received 10.06.2020
Accepted 12.08.2020
