=Paper=
{{Paper
|id=Vol-2164/paper8
|storemode=property
|title=A De-centralized Framework for Data Sharing, Ontology Matching and Distributed Analytics
|pdfUrl=https://ceur-ws.org/Vol-2164/paper8.pdf
|volume=Vol-2164
|authors=Vasileios C. Pezoulas,Konstantina D. Kourou,Themis P. Exarchos,Vassiliki Andronikou,Theodora Varvarigou,Athanasios G. Tzioufas,Salvatore De Vita,Dimitrios I. Fotiadis
|dblpUrl=https://dblp.org/rec/conf/semweb/PezoulasKEAVTVF18
}}
==A De-centralized Framework for Data Sharing, Ontology Matching and Distributed Analytics==
https://ceur-ws.org/Vol-2164/paper8.pdf
A De-centralized Framework for Data Sharing, Ontology
Matching and Distributed Analytics
Vasileios C. Pezoulas1, Konstantina D. Kourou1, Themis P. Exarchos1,2, Vassiliki An-
dronikou3, Theodora Varvarigou3, Athanasios G. Tzioufas4, Salvatore De Vita5, and
Dimitrios I. Fotiadis1
1 Unit of Medical Technology and Intelligent Information Systems, Dept. of Materials Science
and Engineering, University of Ioannina, GR 45110 Ioannina, Greece
{bpezoulas, konstandina.kourou}@gmail.com, fotiadis@cc.uoi.gr
2 Dept. of Informatics, Ionian University, GR 49100 Corfu, Greece
themis.exarchos@gmail.com
3 Division of Communication, Electronic and Information Engineering, School of Electrical and
Computer Engineering, National Technical University of Athens, GR 15780 Athens, Greece
{vandro, dora}@telecom.tuc.gr
4 Dept. of Pathophysiology, Faculty of Medicine, National and Kapodistrian University of Ath-
ens, GR 15772 Athens, Greece
agtzi@med.uoa.gr
5 Clinic of Rheumatology, Dept. of Medical and Biological Sciences, Udine University, IT
33100 Udine, Italy
salvatore.devita@asuiud.sanita.fvg.it
Abstract. In this paper we present the fundamental concept, as well as, the ob-
jectives and the architecture of HarmonicSS EU project. The proposed architec-
ture envisages to address (i) the lack of patient stratification in large patient co-
horts, (ii) the dataset heterogeneity across these cohorts, and (iii) the incomplete
understanding of disease pathogenesis. In order to do so, ontology matching
mechanisms are recruited for harmonizing the clinical data, distributed data ana-
lytics services have been designed for the de-centralized analysis of patient data,
and different health policies can be assessed, all of them with respect to the data
protection guidelines posed by the upcoming General Data Protection Regulation
(GDPR). The HarmonicSS architecture has been employed towards the develop-
ment of a data sharing, harmonization and distributed analytics framework for
the recruitment and analysis of heterogeneous regional, national and international
longitudinal cohorts of patients with primary Sjögren’s syndrome (pSS). The out-
comes can be used by researchers, clinicians, patients, and health policy makers.
Keywords: ontology matching, data protection, distributed data analytics, pri-
mary Sjögren’s Syndrome
1 Introduction
Recent advances in medical data sharing highlight the necessity of large cohort studies
in order to validate the accuracy of existing prediction models for disease prognosis,
2
genetic biomarkers and health policies as well. Moreover, the existing lack of patient
stratification (i) increases the risk of producing incomplete results in clinical trials
which employ highly expensive drugs and (ii) makes difficult the definition of evi-
dence-based health policies. The application of existing or newly identified therapeutic
targets in common clinical practice is hampered by (i) the lack of validation of potential
indices in large patient cohorts, (ii) the dataset heterogeneity across these cohorts, and
(iii) the incomplete understanding of disease pathogenesis. In order to fulfill these
needs, data sharing, data harmonization, and data analytics are necessary.
Here, we present the HarmonicSS platform [1], a de-centralized framework that aims
to address the aforementioned needs. Ethical, legal, and privacy issues for data sharing
among the heterogeneous cohorts are taken into consideration in order to make this
concept possible. Ontology matching tools have been recruited for overcoming the co-
horts’ heterogeneity and distributed learning environments have been adopted for se-
cure processing of the cohort data. The proposed architecture is used to develop a
straightforward data sharing, harmonization and distributed analytics framework for the
analysis of heterogeneous regional, national and international longitudinal cohorts of
patients with primary Sjögren’s syndrome (pSS). Since pSS is relevant not only due to
its clinical impact but also as one of the few diseases to link autoimmunity, and cancer
development, its examination can establish research in many areas of medicine.
2 Towards a federated architecture
The major modules towards the HarmonicSS vision are clearly depicted in Figure 1
along with the users of the platform. These modules include the data sharing, cohort
data harmonization, data sharing management, workflow, and distributed data analyt-
ics, which are further described in the sequel. The HarmonicSS platform will offer a
variety of services, including big data mining tools for lymphoma prediction and patient
stratification, tools for analyzing genetic data, evaluation of health policy scenarios,
social media analytics, a training tool used for educational purposes by clinicians and
patients, a patient selection tool for multinational clinical trials, among others.
The proposed architecture has been adopted towards the harmonization and analysis
of clinical data across 22 cohorts (approx. 7500 patients in total) including a variety of
pSS clinical data, such as saliva, blood and tissue samples, serum, biopsies, tears, DNA,
and RNA samples. In short, pSS is a chronic inflammatory autoimmune disease causing
salivary gland dysfunction (dryness in the eyes, mouth, skin, vagina), affecting primar-
ily women near the menopausal age (almost 10 females per 1 male, followed by sys-
temic sclerosis, systemic lupus erythematosus) [2, 3]. In 40-60% of pSS patients, ex-
traglandular involvement is also exhibited [3], whereas 5% of pSS patients are associ-
ated with the development of B-cell non-Hodgkin lymphoma [2]. These findings sup-
port the potential of several histopathologic, cellular and molecular indices to serve as
biomarkers for the classification of SS patients in distinct groups, the prognosis of dis-
ease severity and lymphoma development, and the selection of appropriate treatment.
Each cohort must fulfill all the necessary data protection requirements prior to the final
upload of the clinical data into the HarmonicSS platform.
3
Fig. 1. The major architectural modules towards the HarmonicSS vision.
2.1 User access management
An OAuth 2 authorization framework has been adopted for user authentication and se-
cure transfer of all the services within the platform based on Secure Sockets Layer
(SSL)/Transport Layer Security (TLS) encryption protocols. Asymmetric key encryp-
tion algorithms have been implemented for the secure transfer of the clinical data into
the private cloud spaces that lie within the cloud. The cloud is the environment where
the HarmonicSS vision takes place. Six Kernel-based Virtual Machines (KVMs) have
been already assigned for the HarmonicSS services.
2.2 Data sharing assessment
Data sharing comprises the core of the proposed federated platform. The data sharing
framework is composed of two major modules, namely the data sharing assessment and
the data sharing management modules. Both of them have been designed to enhance
the secure evaluation and processing of the patient data with respect to the patients’
privacy according to the General Data Protection Regulation (GDPR). The data sharing
assessment module has been designed to ensure that data sharing fulfills all the neces-
sary GDPR requirements from the data provider’s point of view whereas the data shar-
ing management module supervises all the processing procedures that take place within
the platform from the data processor’s point of view. Data providers (controllers) are
responsible for providing all the required documents (e.g., signed informed consent
forms, purposes of processing, transfer to third countries, data protection guarantees,
4
legitimate interests). The latter is conducted by a three-member data controller com-
mittee (DCC) with expertise on the pSS domain knowledge. The clinical data are finally
stored on a private cloud space that is specifically designed for each cohort. Finally, a
data curator mechanism is also provided for outlier detection and removal, de-duplica-
tion, attribute identification and attribute grouping, missing values detection and auto-
matic fill based on a variety of data imputation methods.
2.3 De-centralized data harmonization
Data harmonization aims to overcome the heterogeneity of medical cohorts worldwide
by converting the heterogeneous datasets into compatible ones with minimum loss.
Harmonization involves several mechanisms including, cohort data transformation, ter-
minology description, and ontology alignment. The main idea behind medical data har-
monization is based on the mapping of each template of interest into a (pre-defined)
reference template (model). The HarmonicSS reference model has been co-designed
with the assistance of clinicians to meet the minimum requirements for effectively de-
scribing the pSS domain knowledge. Protégé [4] has been used to convert this reference
model to a database ontology which comprises the main ontology. Semantic interlink-
ing mechanisms are then used to extract the structure and the vocabulary of each source
dataset for enabling ontology matching. Ontology matching is conducted by mapping
the ontology of a source dataset into the main ontology (which serves as the target on-
tology) using semantic matching tools [5, 6]. The semantic matching algorithm pro-
vides all possible relational associations between each term of the source ontology with
those from the target ontology. Then, the most appropriate terms are selected according
to clinical guidelines. Novel software tools that are often employed for data harmoni-
zation include the SORTA tool [6] for data re-coding and annotation, the S-Match tool
[7] for overcoming the semantic interoperability problem, the BiobankConnect soft-
ware [8] for ontological and lexical indexing, and the Opal software which has been
developed under the EU BioSHaRE project [9]. The harmonized data will be finally
stored on the private cloud spaces of the corresponding cohorts.
2.4 Data sharing management
A data processor who wishes to process one or more de-centralized clinical cohort data
(e.g., evaluate an existing or a new lymphoma prediction model on different cohorts)
must first request access to the private cloud space of the corresponding cohort(s). The
data providers that manage these private repositories can either provide the green light
for allowing the local processing (handshaking) or not. If so, the services are executed
locally on the private cloud spaces and the results are combined according to the dis-
tributed learning concept based on which the clinical data never leave the hospital (clin-
ical center). This enhances the secure processing of the clinical data but often introduces
biases during the analysis due to the heterogeneity of the cohort data structures. Har-
monization is a promising solution that overcomes this heterogeneity as mentioned be-
fore. The data providers are also responsible for providing regular reports about their
data management status (e.g., any actions performed by the data processors on the pa-
tients’ data). Notifications can also be sent between the users of the platform.
5
2.5 Workflow
The workflow module is the basis for executing the data analytics services. Workflows
are organized in a specific manner which includes input(s), processing stages and out-
put(s). A workflow mechanism consists of the workflow builder which allows the user
to build a model, the workflow loader which allows the user to load the constructed
model and the workflow manager which allows the user to edit, execute or delete ex-
isting workflows. The user can access the workflow mechanism through the web client.
For example, the workflow module allows the user to upload an existing prediction
model and evaluate it on the HarmonicSS clinical cohort data.
2.6 Distributed data analytics
Data analytics consists of a variety of services including genetic data analytics, big data
mining services, health policies impact assessment, social media analytics, patient se-
lection tool for clinical trials, training tool for educational purposes. Genetic data ana-
lytics involves the identification or confirmation of existing SNPs, correlation of exist-
ing SNPS with disease sub-phenotypes, case-control and case-case associations, among
others. The big data mining services consist of data mining algorithms for constructing
data models (e.g., for lymphomagenesis prediction) which can be either executed on an
individual cohort’s private cloud space (one site analysis using classic algorithms, such
as, support vector machines, logistic regression, decision trees, neural networks, etc.)
or on multiple private cloud spaces using distributed data mining approaches (multisite
analysis) [10] for security purposes. The health policies impact assessment procedure
involves the creation, modeling, and assessment of new health policy scenarios. These
models can be used to propose a drug to be tested on an appropriately selected patient
and improve the patient diagnosis and screening procedure.
3 Discussion
The HarmonicSS infrastructure follows European guidelines for the management of
pSS patients in order to derive rules for storing blood, tissue, saliva, serum, DNA, RNA
and biopsies samples in biobanks for pSS. Prior to the data harmonization procedure, it
is important to identify missing information from the cohorts. HarmonicSS has already
defined a set of minimum criteria necessary at diagnosis/follow-up including pathologic
ocular involvement disease indicators, such as, Schrimer’s, and laboratory measures,
such as, leukopenia, serum, cryoglobulinemia, tears, lip or parotid biopsy, ESSPRI and
ESSDAI scores, etc., for improving data inclusion and quality as well.
Since the HarmonicSS cohorts contain approximately 500 patients that have pro-
gressed to lymphomagenesis, the risk prediction of lymphoma development will be ef-
fectively addressed. Correlation analysis between harmonized genetic phenotypes can
possibly lead to the identification of new biomarkers and/or validation of existing ones,
with greater accuracy. Since no data pooling is performed but rather a decentralized
approach for harmonization and analytics is adopted, the security of the patient data is
well-preserved and thus the platform’s reliability and impact is greatly enhanced. The
6
HarmonicSS ‘data protection by design’ (GDPR compliant) architecture along with the
distributed data processing services through appropriate requests, comprise a novel
strategy for the construction of an innovative federated health platform for pSS patients
which further considers for biobank creation and maintenance as well.
The HarmonicSS framework details a novel methodology for the initialization,
maintenance, and expansion of heterogeneous data infrastructures. The overall im-
portance and major contribution of HarmonicSS lies in the fact that it is performed on
a cross-country data infrastructure of unique heterogeneity and content variability.
Great emphasis is given on the development of semantic interlinking tools for homog-
enizing the clinical data, as well as, on the establishment of distributed environments
for analyzing the harmonized data. Results linking evidence to practice can be used to
support policy makers, and health management researchers to define novel health rec-
ommendations for pSS prevention to be adapted to a shared health policy environment.
Acknowledgement. This project has received funding from the European Union’s
Horizon 2020 research and innovation programme under grant agreement No 731944
and from the Swiss State Secretariat for Education, Research and Innovation SERI un-
der grant agreement 16.0210
References
1. HarmonicSS: HARMONIzation and integrative analysis of regional, national and interna-
tional Cohorts on primary Sjögren’s Syndrome (pSS) towards improved stratification, treat-
ment and health policy making, http://harmonicss.eu/.
2. Mavragani, C, and Moutsopoulos, H.: Sjögren syndrome. CMAJ, 186(15), E579-586 (2014).
3. Ramon-Casals, M., Brito-Zerón, P., Sisó-Almirall, A., and Tzioufas, A. G: Topical and sys-
temic medications for the treatment of primary Sjögren’s syndrome. Nat. Rev. Rheum., 8,
399-411 (2012).
4. Protégé: a free, open-source ontology editor and framework for building intelligent systems,
http://protege.stanford.edu/
5. Christophides, V., Efthymiou, V., and Stefanidis, K.: Entity resolution in the web of data.
Synthesis Lectures on the Semantic Web, 5(3), 1-122 (2015).
6. Pang, C., Sollie, A., Sijtsma, A., Hendriksen, D., Charbon, B., de Haan, M., de Boer, T.,
Kelpin, F., Jetten, J., and van der Velde, J. K.: SORTA: a system for ontology-based re-
coding and technical annotation of biomedical phenotype data. Database, 2015 (2015).
7. Giunchiglia, F., Autayeu, A., and Pane, J.: S-Match: an open source framework for matching
lightweight ontologies. Semantic Web, 3, 307-317 (2012).
8. Pang, C., Hendriksen, D., Dijkstra, M., van der Velde, K. J., Kuiper, J., Hillege, H. L., and
Swertz, M. A.: BiobankConnect: software to rapidly connect data elements for pooled anal-
ysis across biobanks using ontological and lexical indexing. J. Amer. Med. Assoc., 22, 65-
75 (2014).
9. Doiron, D., Burton, P., Marcon, Y., Gaye, A., Wolffenbuttel, B. H. R., Perola, M., Stolk, R.
P., Foco, L., Minelli, C., Waldenberger, M., Holle, R., Kvaloy, K., Hillege, H. L., Tasse, A.
M., Ferretti, V., and Fortier, I.: Data harmonization and federated analysis of population-
based studies: the BioSHaRE project. Emerg. Themes Epidemiol., 10, 12 (2013).
10. Tsoumakas, G., and Vlahavas, I.: Distributed data mining. Encyclopedia of data warehous-
ing and mining (2009).