gms | German Medical Science

Introduction: (Fast Healthcare Interoperability Resources) is a modern standard for communication and representation of clinical data, where each dataset is defined by a single resource, linked to other resources [1]. But FHIR^® does explicitely not define how to persist these resources [2]. Most implementations of a persistence-layer for FHIR^® resources are using a relational database management system (RDBMS) [3], [4], [5].

Since a relational database require queries to be conducted on tables or views, partly using expensive JOIN functions, and filtered for the exact dataset, they tend to become slow on larger databases and may require the RDBMS to cache huge amounts of the stored data during the query [6].

Within the last years we've seen a rise of Graph Database systems for the storage of large interlinked datasets. Those database systems may pose a perfect match for a persistence layer of FHIR resources. Especially in clinical studies where the data is equally important as its relations [7].

Methods: We conducted basic research on how to implement a graph data model for persistence of FHIR^® resources. This model shall be used for decision making on a future implementation of such a persistence layer. Since the resulting model is important for the determination of a database system, the result shall be technology agnostic.

For an initial concept of the model we decided to constrain the resources to a subset to fulfill the basic needs of a short inpatient stay with multiple Observations and Procedures. The model was designed through a review of these resources and a structured definition of nodes, edges and properties in a property graph model [7], [8].

Results: Our resulting data model shows, that a graph data model is well suited for the data structure of interlinked FHIR resources. Also the representation in a graph model renders very comprehensible.

The graph model includes a node for each FHIR^® resource, containing properties for each of the resources elements and an edge for each "Reference" element within a resource. In order to maintain data provenance, each node itself begins a directed acyclic graph containing newer or older versions of this resource as nodes linked via edges with a specific label.

Discussion: At first glance the structure of a graph data model seems well suited for our purpose. Nevertheless the model still has to prove its usefulness. Due to no real implementation, we did not conduct any queries or other real tests on the model.

We expect some queries to be substantially easier and faster than in classic RDBMS. Especially queries on relations to a specific node (i.e. "Observations of Patient xyz"), but we need to evaluate extensive queries for specialized data sets. Especially regarding the common query for specific clinical resources over a given time frame (i.e. “number of infections in the last year”) and further restrictions (i.e. “... within ICU”).

In order to test the model and do further research we will need to implement this model in a graph database system.

The authors declare that they have no competing interests.

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