gms | German Medical Science

Introduction: To answer research question considering medication risks in routine health care applicable at several German university hospitals, we had to develop a distributed analysis approach, the “POLAR_MI ETL Pipeline”, from the scratch.

State of the art: POLAR_MI is one of the first projects in Germany taking advantage of the newly established data integration centres (DIC). Hence, the knowledge in implementing distributed approaches was rather limited. One goal of the DIC is to create an overarching and comparable level of data presentation following the specification of the internationally balloted MII core data set (CDS) including the definition of required FHIR profiles.

Concept: The POLAR_MI researchers decided to establish a recurring release and development cycle. In a first step, pharmacists and physicians discussed the research questions that should be answered. Then, retrieval and statistical analysis scripts (programmed without direct data access) were made available to the collaborating DIC. After script execution, the DIC provided a feedback to the POLAR_MI team. Based on this, the research questions were refined, the scripts adapted and the development cycle restarted.

Implementation: As the DIC had to provide the clinical data on FHIR servers, the POLAR_MI framework had to handle FHIR queries while being executed at the local hospital infrastructure. Hereby we took advantage of the R-package fhircrackr developed in our team [1] that can be used to retrieve data from FHIR resources. The framework itself, provided as docker container, was implemented as a two-step distributed approach, consisting of a local data retrieval and analysis part at each DIC, followed by a central meta-analysis. This procedure took care that any confident (personal) clinical data remained locally. The local part finished with an aggregation of information that was finally submitted to a cloud structure.

Lessons learned: The landscape of German university hospital settings and local IT infrastructures raised various issues. Starting with the expectation to analyse harmonized data from various FHIR endpoints, we faced problems that we assumed they had already been solved. This comprises differences arising with implementation of various FHIR server technologies such as Blaze, HAPI or IBM, the server resources available with regard to CPU, RAM or storage, the handling of missing data or the interpretation of CDS specification. Hence it was impossible to identify a setup enabling a robust local retrieval and analysis process for each collaboration partner equally. Furthermore, during the development and release cycle we had to deal with issues resulting from initially imprecise specifications within the CDS. For instance, it is mandatory to provide units related to lab values, whereas the unit itself could be stored as SI unit, human or machine-readable form. In addition, the FHIR architecture itself did not appear fully developed. Even for the same state of data elements, the FHIR server responded differently to identical queries without recognizable reason.

The interdisciplinarity of the POLAR_MI team was an advantage during this challenge to identify the root of such issues including script errors, (mis-)interpretation of CDS specifications or simply an information loss during data transfer within the hospital.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.