The development and deployment of electronic health records (EHRs) worldwide has led to several interoperability challenges, including extracting valuable information from free-text clinical notes, integrating unstructured EHR patient data, and reusing them without ambiguity for clinical research purposes [1,2].

Interoperability is defined by the Institute of Electrical and Electronics Engineers (IEEE) as the “ability of two or more components to exchange information and to use the information that has been exchanged” [3]. A total of 4 interoperability levels were proposed by the Healthcare Information and Management Systems Society organization: (1) foundational, (2) structural, (3) semantic, and (4) organizational (Figure 1) [4].

Important research efforts are underway to develop health care standards, including Fast Healthcare Interoperability Resources (FHIR) standards for the electronic exchange of health care information. This standard was developed by the Health Level Seven International (HL7) organization aiming at the evolution of messaging standards to achieve semantic interoperability [5]. FHIR combines the best features of HL7 version 2, version 3, and clinical document architecture, while leveraging the latest web standards.

FHIR interoperability solutions are designed from a set of modular components called “resources” that have been designed as small reusable components defining a set of properties for the capture of the structure of the domain data acquisition for a particular use. The version 4.3.0 release of FHIR R4B proposes 143 different “resources” classified into five categories to handle health data: (1) foundation resources, (2) base resources, (3) clinical resources, (4) financial resources, and (5) specialized resources (Multimedia Appendix 1).

Many elements in FHIR use coded values and allow the use of external terminologies, such as (1) Systematized Nomenclature of Medicine–clinical terms (SNOMED CT)—comprehensive multilingual clinical terminology used in EHR. Its components include concepts (codes), descriptions (terms), and relationships—and (2) Logical Observation Identifiers Names and Codes (LOINC), a set of universal names and ID codes used to identify laboratory and clinical test results.

Common terminology aims to address the problem of ambiguity in exchanged data [6]. However, most legacy systems do not take advantage of a terminology-based approach to data exchange and still use incompatible terminologies or unstructured data in their exchanges with EHR. Thus, to ensure semantic interoperability with these legacy applications, it is mandatory to perform manual data transformations, mapping, or other alternate solutions to ensure that the exchanged data are interpreted correctly by the information systems involved.

The FHIR standard is of great interest to researchers and EHR ecosystem developers [7], and several technology giants, such as Amazon, Google, IBM, Microsoft, Oracle, and Salesforce, offer ready-to-use cloud services to promote the adoption of FHIR [8].

In this study, we report on a systematic mapping Review (SMR) and propose a classification of the FHIR literature based on FHIR resources, terminologies, research approaches, and innovative contributions. The goal of this SMR was to provide state of the art, trends, and analysis of research related to semantic interoperability using FHIR. This activity specifically aimed to identify the FHIR and terminology used, as well as the different proposed approaches, contribution types, and trends.

Of the number of literature reviews on the challenges of health information systems interoperability, only 4 specifically address FHIR standard issues [2,9-11]. This section summarizes the findings of this study.

The study by Ayaz et al [10] reviewed 80 studies published between 2012 and 2019 that identified the opportunities and challenges associated with using FHIR. Another FHIR systematic review by Lehne et al [9] identified the main issues to be considered when exchanging data based on only 2 data sources and presented an overview of 131 selected studies up to February 4, 2019. A third systematic literature review by Setyawan et al [2] on data interoperability issues and possible solutions selected 39 studies published between 2018 and 2021 from 3 data sources (ie, IEEE Xplore, ScienceDirect, and Scopus). Finally, a systematic review of semantic interoperability in the study by de Mello et al [11] selected 28 studies from 2012 to 2020 from 8 different sources.

An overview of these 4 studies is presented in Table 1 in terms of their study goal or objective, period covered, number of studies selected and analyzed, source databases, and their respective mapping and research questions (RQs).

To the best of our knowledge, no systematic mapping review has been published that specifically addresses the semantic issues raised when using FHIR.

Some of the lessons learned in these studies and their limitations are presented in Table 2. The key findings are summarized as follows:

These previous studies do not provide an overview of post-2019 developments related to semantic interoperability approaches and trends using FHIR. An extensive and more recent study is needed using many databases and a search string that includes the most possible semantic keywords.

In contrast to related works, this SMR focuses on the semantic level of interoperability using FHIR and reports on research directions over the last decade.

SMRs and systematic literature reviews typically use the same methodology, but their goals differ. The systematic literature review is driven by specific questions and the review aims to respond to these specific questions. In contrast, an SMR targets more high-level questions aimed at a specific topic, with the goal of providing a classification of the research results. An SMR can be helpful to researchers because it provides an overview of the literature at a given time for a specific topic of interest [12].

According to the study by Petersen et al [13], the mapping review process follows three steps: (1) planning the mapping, (2) conducting the mapping, and (3) reporting the results of the mapping. Furthermore, specific guidelines for a systematic mapping review in the software engineering domain as mentioned in the study by Brereton et al [14] suggest more detailed activities at each step (Figure 2):

A set of 5 questions was defined (Table 3).

QA of a research publication is a crucial selection step to ensure that the selected studies have the appropriate impact for mapping reviews and further analysis [12]. For this purpose, we defined quality criteria scores based on the following questions.

Quality evaluation of the identified and extracted studies was performed by answering each QA question. The selected studies satisfied at least 4 of the 6 QA criteria. Others were rejected.

This SMR was conducted in 2022, with publications published from 2012 to 2022. A summary of the data extraction showing the selection filter effect is presented in Figure 3.

The 70 studies retained are listed in Multimedia Appendix 2 [16-85]. The extracted data are presented in Multimedia Appendix 3.

We reviewed 70 selected studies and identified the types of FHIR resources that were used. Table 4 lists the 62 types of resources identified in the corresponding studies that have used them. In Figure 4, the resource types are presented, from most frequently used to least used.

The 10 most referred or used resource types in these studies were Observation (31/297, 10.4%), Patient (29/297, 9.8%), Medication (19/297, 6.4%), Condition (16/297, 5.4%), CodeSystem (15/297, 5.1%), Encounter (11/297, 3.7%), Procedure (11/297, 3.7%), Medication Request (10/297, 3.4%), Practitionner (10/297, 3.4%), and ValueSet (9/297, 3%).

The HL7 organization classifies the list of resources into five categories: (1) foundation, (2) base, (3) clinical, (4) financial, and (5) specialized resources. Each of these categories was further divided into 5 subcategories, except for the financial resources category, which was subdivided into 4 subcategories.

The resources inventoried in the 70 selected studies were mostly in 3 clinical subcategories: summary, medications, and diagnosis (Figure 5 and Multimedia Appendix 1). There is one exception to Patient and Encounter resources, which are in the base category and mainly used for identification purposes. We also noted the terminology subcategory (foundation category) among the top 10 most used resources. Terminology resources were used to map concepts to FHIR terminology (eg, SNOMED CT and LOINC).

A total of 9 distinct FHIR terminologies were used in the 70 selected studies (Table 5).

Their distribution and percentage are as follows:

A total of six main semantic research approaches were identified from the solutions proposed in the 70 selected studies:

Table 6 presents the semantic research approaches used to decrease the frequency of use in the selected studies.

The distribution of the 6 semantic approaches is as follow: mapping (31/126, 24.6%), RDF or OWL (24/126, 19%), ML and NLP (20/126, 15.9%), annotation (18/126, 14.3%), terminology (18/126, 14.3%), and ontology (15/126, 11.9%).

The selected studies used at least one of the 6 identified semantic research approaches, and most studies used more than one research approach. For example, 22 of the 31 studies listed in Table 6 used mapping and other approaches (70%), whereas only 29% (9/31) used mapping alone.

To better understand the frequency of combinations of more than one semantic research approach, we identified a number of studies using more than one approach. We performed the same exercise for RDF or OWL followed by annotation. Table 7 shows that most combined semantic approaches are annotation and ML and NLP, mapping and terminology services, mapping and annotation, RDF or OWL, and ontology.

The next analysis related to the proposed approach examined how these approaches evolved over time. Figure 6 shows how these approaches were used over time. The main finding are (1) ML and NLP-based proposals emerged in 2018 in FHIR semantic interoperability studies, (2) the other approaches span the full period, and (3) no specific research approach dominates the FHIR semantic interoperability research field.

The objective of RQ4 was to identify the contribution types within each of the 70 selected studies. For this purpose, five main contributions were identified: (1) new framework or architecture intended to enable semantic interoperability (framework or architecture proposal), (2) evaluation or comparison of different techniques (eg, comparison of ontologies), (3) new ontology, model or enhancement of an existing one (model proposal), (4) a new technique or improvement of an existing technique (new technique proposal), and (5) a tool proposal through the development of a new tool or an enhancement to existing ones to facilitate achieving semantic interoperability (tool proposal).

Table 8 presents the distribution of the 70 studies selected according to contribution type.

Figure 7 shows the distribution of studies by contribution type.

The most frequent contribution type was the proposal of a framework or architecture (26/70, 37%), followed by model proposals (13/70, 19%), new or improved techniques (13/70, 19%), comparison services (10/70, 14%), and tool proposals (8/70, 11%).

This section presents a discussion of each of the key findings of the RQs.

This study was limited to semantic studies using only the FHIR standard, and it did not search for other studies related to semantic interoperability in general; therefore, this SMR focused exclusively on studies addressing FHIR implementation.

Furthermore, only studies from the 10 database sources specified were considered, and commercial solutions were excluded.

During the classification of the contribution based on 5 categories, 2 studies (S40 and S49) were considered in 2 different categories (model and technique). Both situations were evaluated and the results were only slightly different.

This SMR inventoried and analyzed the findings on semantic interoperability using FHIR over the last decade (2012-2022) in 70 selected studies from 10 databases. It presents a classification of the selected studies according to the semantic approaches used and their contributions. It also provides the frequency of use of FHIR resources and terminologies.

These findings provide researchers with an inventory of the approaches used to achieve semantic interoperability using the FHIR. The SMR also documented the frequency with which the proposed approaches were used. This may help identify avenues for research on semantic interoperability. It can also be used as a guideline by the FHIR community for the future development of approaches and types of resources that are not well covered in research published to date.

Indeed, understanding and interpreting the exchanged data in the EHR ecosystem needs to be supported by further research using, for instance, FHIR in mapping, RDF or OWL, and comparing ontologies and terminology services.

Semantic interoperability is still an active research field, and there is no broad consensus in the health care community on how to achieve full semantic interoperability between information systems. The main issues reported are related to the difficulty associated with the use of different terminologies and the efforts required to successfully map to FHIR terminology. Even worse, when no terminology is used, unstructured data would still need to be extracted and mapped to the appropriate concept code for FHIR terminology.

This SMR provides a classification of the different solutions proposed to address semantic interoperability using FHIR at different steps: collecting, extracting and annotating data, modeling EHR data from legacy systems, and applying transformation and mapping to FHIR models and terminologies. The use of ML and NLP for unstructured data is promising and has been applied to specific use case scenarios. In addition, terminology services are needed to accelerate their use and adoption; furthermore, techniques and tools to automate annotation and ontology comparison should help reduce human interaction.

Future work will be devoted to the FHIR-based framework for the semantic interoperability of EHR data.