-
Problem statement: A growing volume and variety of personal health
data are being collected by di erent entities, such as healthcare providers,
insurance companies, and wearable device manufacturers. Combining heterogeneous
health data o ers unprecedented opportunities to augment our understanding
of human health and disease. However, a major challenge to research lies in
the di culty of accessing and analyzing health data that are dispersed in their
format (e.g. CSV, XML), sources (e.g., medical records, laboratory data),
representation (unstructured, structured), and governance (e.g., data collection and
maintenance)[
2
]. Such considerations are crucial when we link and use personal
health data across multiple legal entities with di erent data governance and
privacy concerns.
Proposed approach: Our approach to tackling this challenge is extending
the Personal Health Train Architecture [
1, 2
] to analyze health data from
multiple sources in a privacy-preserving manner[
5
]. Instead of centralizing the data
for the analysis, researchers send data-processing algorithms (application trains)
to each data source. To be able to run the application train, each source deploys
a data station which stores the data required for the analysis. The data station
only returns the results of the analysis rather than any of the original data.
Data stations should be able to provide data in certain standard formats and
structures in line with the FAIR (Findable, Accessible, Interoperable, Reusable)
principles[
3
]. In our method, the FAIR data stations are based on Semantic
Web technologies. The data is stored as a knowledge graph using the Resource
Description Framework (RDF), while its representation complies with the
ontologies accepted by the community. We apply the Data2Service framework[
4
],
rst described at the 2018 SWAT4LS, to now semi-automatically transform and
integrate heterogeneous health data into RDF data. This knowledge graph
complies to a target set of ontologies, and is subsequently made available as FAIR
data stations through a set of interfaces and services (e.g., SPARQL, API). The
framework de nes a set of scalable and sustainable transformation work ows
to convert any structured data sources to a target data model. To help the
researchers with the mapping process, Data2Services generates SPARQL mappings
les based on the input data structure. Then, researchers create the analysis as
an application train to query the RDF data available at FAIR data stations
using SPARQL, pre-process data, and execute machine learning models[
6
]. The
FAIR data stations are always in the governance of the data entities. Instead of
duplicating the original data and sending to the researchers, our method ensures
the original data will not leave the data entities and be exposed to any others
including the researchers. The researchers only receive the results of applications.
Use case: The goal is to study annual healthcare costs in relation to the
incidence of Type 2 Diabetes Mellitus (T2D) without revealing any original data.
We used patients' health data from De Maastricht Studie, which is a
population study on T2D, and their healthcare cost data from Statistics Netherlands.
The original data les (SAV, CSV) were automatically transformed to RDF by
de ning SPARQL construct queries, and subsequently loaded into a triple store
and made available as per the FAIR data station speci cation. Researchers can
retrieve RDF data by SPARQL queries from FAIR data station and generate
to CSV, JSON, XML formats for the data analysis. The variable names, labels,
keywords were used to nd the most relevant ontologies using the NCBO
Ontology Recommender[
7
]. RDF data was structured to the relevant ontologies
Ontology of Consumer Health Vocabulary and SNOMED Clinical Terms. The
structured RDF data is stored at FAIR data stations, with metadata compliant
to the HCLS Dataset Description Pro le. Analysis models such as correlation
matrix, linear regression were successfully tested in the infrastructure.