A comparison of Fast Healthcare Interoperability Resources and Observational Medical Mutcomes Partnership electronic health record data within the All of Us Research Program

Abstract

Objective

This study compares the contents of two data standards; the Observational Medical Outcomes Partnership (OMOP) and Fast Healthcare Interoperability Resources (FHIR), highlighting their strength and weaknesses and serve as an initial step toward understanding how each standard supports secondary data analysis.

Materials and Methods

Participant electronic health record data in both OMOP and FHIR formats from the All of Us Research Program (AoURP) were compared, including codeable event volume, healthcare encounters, and person timelines. A phenotype-based assessment was also conducted using Type-II Diabetes Mellitus (T2DM).

Results

Among 29 512 participants identified with overlapping FHIR and OMOP data, Median codeable event counts were comparable between FHIR and OMOP within the Measurement (OMOP = 846; FHIR = 832), Drug (OMOP = 92; FHIR = 90), and Observation (OMOP = 65; FHIR = 100) domains, but were higher in OMOP within the Condition (OMOP = 258; FHIR = 11) and Procedure (OMOP = 72; FHIR = 4) domains. Within the T2DM cohort, OMOP contained more data, except for medications. Very few participants had encounters in FHIR (1.5%) relative to OMOP (97.0%). On average, only 15.9% of visit dates overlapped in both standards, with most visit dates occurring only in OMOP (65.3%) or only FHIR (18.4%).

Discussion

Data in OMOP showed a higher volume of observation and procedure codeable events, reported encounters, and T2DM symptoms, complications, and comorbidities. FHIR data was able to capture data across multiple providers and health systems.

Conclusion

Based on AoURP data, both standards were shown to support healthcare data capture, although OMOP shows greater utility for research purposes as its extract, transform, and load process enables more flexible data capture relative to extracting data from FHIR payloads. FHIR, however, captures patient-level data from beyond the health system and can thus be used to supplement OMOP.

Background and significance

Electronic health record (EHR) data is increasingly used to conduct research.^1,2 Such data is collected during the delivery of care and contains extensive information about disease diagnosis, progression, and treatment. Recently, several data networks have formed with the aim of improving EHR-based study generalizability³ by facilitating multi-site research.^4–6 Consolidating this data, however, is a major challenge, both on syntactic (connecting and exchanging information) and semantic (interpreting and using information effectively) levels.⁴

In response to this challenge, several common data models (CDMs) have been developed to standardize clinical data using shared terminologies.⁷ Two widely used data standards are the Observational Medical Outcomes Partnership (OMOP) and the Fast Healthcare Interoperability Resources (FHIR).⁸ OMOP is an open-source relational data model developed to support post-market drug safety surveillance.⁹ OMOP is currently maintained by the Observational Health Data Sciences and Informatics (OHDSI) research community, who have developed a suite of analytic tools designed to interrogate OMOP data.¹⁰ OMOP has consistently been shown to be well-suited for EHR-based observational research, as reflected by its widespread adoption across large research networks,^11–13 including the All of Us Research Program (AoURP)⁵ and the National Clinical Cohort Collaborative.⁴

FHIR was developed by Health Level Seven (HL7),¹⁴ a global authority on healthcare interoperability data standards. The first FHIR standard was released in 2014 with the aim of improving interoperability between healthcare systems and applications. FHIR was designed to be a flexible, scalable, and easy to implement data standard, which led to its widespread adoption. Legislative mandates, like the 21st Century Cures Act (2016),¹⁵ further accelerated this adoption by requiring certified health IT developers and EHR vendors to implement FHIR.^15,16 Concurrently, the Office of the National Coordinator for Health Information Technology¹⁷ has promoted the United States Core Data for Interoperability (USCDI) to ensure a minimum dataset for FHIR exchange.^8,18

While both OMOP and FHIR organize data into clinical domains and define connections between them, there are several key differences between the two standards. FHIR was designed to securely exchange data between healthcare systems,⁸ whereas OMOP was designed to support observational research. As such, OMOP does not support data exchange between sites, but instead enables researchers to share standardized analysis code that can be run across a federated network of OMOP databases.¹⁹ Additionally, the process of transforming raw data into each standard differs. OMOP requires an extract, transform, and load (ETL) pipeline to map locally coded EHR data to the OMOP standardized vocabularies.²⁰ Conversely, FHIR is a feature within most EHR systems, allowing data entered into the EHR to automatically be converted into a FHIR resource and accessed via a RESTful API²¹ that enables sharing between providers.^22–24 Thus, while FHIR data undergoes changes to be converted from raw data to FHIR, transactional details are handled within the EHR; whereas, transformation to OMOP requires an institution-level ETL pipeline and project-specific transfer mechanism.¹⁹ This process enables all historical data within an institution’s clinical data warehouse to be converted to OMOP, while FHIR only contains data stored in the EHR, which may be lost if the institution changes EHR vendors.

Though each standard was created for different purposes, when used together FHIR and OMOP could be complementary.⁸ Since FHIR was designed for interoperability, it could potentially augment institutional OMOP data by pulling in encounters, labs, medications, and device data that originate beyond individual health systems. Such data could then be harmonized with row-level OMOP data through ETL pipelines designed to transform FHIR payloads into OMOP. This process has already been implemented^25,26 to support tandem use of OMOP and FHIR data for long-COVID research²⁷ and clinical support tools.^28,29

Objective

Given that OMOP and FHIR could be used in tandem, it is important to understand the intricacies of each standard as there is likely considerable overlap but also variation between OMOP and FHIR data. Understanding these intricacies is increasingly important as they could impact the data available in each standard, which could have profound implications for analyses conducted using OMOP and FHIR. Despite this, while the ease of adoption between these two standards has been compared,³⁰ the similarities and differences between OMOP and FHIR’s ability to support observational research have never been characterized. We aim to address this limitation by conducting a comparison of OMOP and FHIR data for a given set of participants within the AoURP consented participant cohort, specifically evaluating each standard’s data volume, data quality (DQ), and suitability for clinical research. Findings from our evaluation are an initial step toward helping researchers and data networks understand how each standard supports different aspects of secondary use of EHR data and how they may be used complementarily.

Materials and methods

Data source

The AoURP is a nationwide research consortium comprised of 50 Healthcare Provider Organizations (HPOs)⁵ who collect and submit data from consenting participants. HPOs submit data quarterly, which includes EHR data transformed to OMOP version 5.3.1 (Figure 1A).³¹ A third-party data aggregator also submits participant-mediated FHIR data, version R4.³² Participants leverage an AoURP study app to connect their AoURP account with hundreds of FHIR providers, providing access to data from multiple FHIR sources per participant. These data are then converted into OMOP by a data aggregator (Figure 1B). FHIR data can contain data from HPO participants, meaning participants can have data available in both standards.

Figure 1.

Clinical data ingestion pipelines. (A) Data ingestion pipeline for OMOP data. AoURP participants are paired with HPO that own internal EHRs. Internally, HPOs execute ETLs to transform their clinical data into the OMOP CDM format, and submit the data to AoURP on behalf of the participant. A feedback-submission cycle then follows between AoURP and the HPO until the data is fully compliant. (B) Data ingestion pipeline for FHIR data. AoURP participants connect to FHIR providers via digital participant portals. These portals extract data via API from clinical datasets in the FHIR format. The data aggregator moderately standardizes the data and submits it to AoURP. No feedback cycle follows.

For the current study, we restricted our analysis to participants aged 18 or older, with at least one record in OMOP and one record in FHIR between January 2020 and March 2024. All analyses were performed on the FedRAMP environment within AoURP’s Data and Research Center’s Google Cloud Platform. All analyses performed were approved by the AoURP Institutional Review Board.

Data comparison

Data harmonization

To enable comparison between the two standards, we aligned OMOP tables and corresponding FHIR resources along a set of clinical domains—persons, visits, conditions, drugs, measurements, procedures, and observations—and conducted parallel analyses across both datasets. The selected domains correspond to core OMOP tables commonly used in observational research. The domains were chosen to enable a direct, like-for-like comparison between OMOP and FHIR in areas where concepts overlap.^11–13 Specifically, we aligned the person, visit_occurrence, condition_occurrence, drug_exposure, measurement, procedure_occurrence, and observation tables within OMOP to 22 related FHIR resources (see Supplementary Materials for mappings).³²

Using these aligned structures, we performed analyses to compare data volume, distribution, and content across OMOP and FHIR. Analyses focused on identifying similarities and differences of select variables in demographic data, healthcare encounters, and codeable event volume in the context of measurement completeness.

Demographics

Chi-square and student’s t-tests were applied to assess if distributions of age, biological sex, race, and ethnicity differed significantly between OMOP and FHIR. Since participants could have FHIR data from multiple sources, participants with consistent values across sources were kept, while participants with conflicting values were excluded from this analysis.

Healthcare encounters

We compared healthcare encounters using three metrics. The first was a per-participant measure to assess overlap of encounters between OMOP and FHIR. We call this metric visit-days and define it as the unique set of days when a participant starts a visit. It is measured using the volume of visit-days that overlap or are disjoint between the two standards. This approach was used to account for multiple encounters occurring on the same day. The second metric was a breakdown of visit types and the third was the distribution of inpatient visit duration.

Codeable event volume

We compared counts and distributions of codable events—defined as a clinical event, such as a diagnosis, procedure, or adverse event, that can be assigned a standardized code from a medical terminology.³³ Due to the nested structure of FHIR, we linked codes embedded within CodeableConcept type fields in FHIR to the OMOP vocabulary (see Supplementary Materials). Using these mappings, we identified codable events from the Condition, Drug, Measurement, Procedure, and Observation tables in OMOP and compared counts using equivalent events across the 22 FHIR resources. Codeable event volume was measured using the number of rows/entries in a domain table/resource. Using these counts, we calculated the median and range of FHIR events per participant per domain and compared them to OMOP.

Timeline comparison

To assess if FHIR can augment single-site OMOP data, we evaluated the contribution of FHIR data to OMOP by identifying participants with at least 10 records in both OMOP and FHIR. For each participant, we calculated the proportion of records that were unique to OMOP, unique to FHIR, or present in both (defined as identical record start dates) and averaged percentages across all participants.

Data quality

We conducted a DQ assessment on both data standards. OMOP was evaluated using OHDSI’s Data Quality Dashboard (DQD),^34,35 while FHIR was evaluated using Google’s FHIR DBT Utils project.³⁶ The DQD applies a set of DQ metrics which are categorized using Kahn’s framework³⁷ to assess data completeness, conformance, and plausibility of EHR data. Completeness measures the presence or absence of data. Conformance measures whether values meet syntactic or structural constraints. Plausibility measures the believability or truthfulness of data values.³⁷ While FHIR DBT Utils evaluates DQ across five categories, our assessment only included metrics from the data completeness and referential integrity categories that were equivalent to DQD metrics (see Supplementary Materials).

Metrics from each tool were harmonized into a set of SQL queries that were run on data in both standards. Metric rates were normalized between 0 and 1, where 1 indicates a perfect score. Row counts for each metric’s DQ requirement were used to calculate proportions and two-sample t-tests of proportions were applied.

Phenotype analysis

To compare fitness of use in clinical analyses, we created a Type II Diabetes Mellitus (T2DM) cohort³⁸ in both OMOP and FHIR. Participants were identified using an adapted version of the T2DM phenotype—used for DQ checks by AoURP—from the Phenotype Knowledgebase (PheKB), a database of well-validated electronic phenotype definitions.³⁹ PheKB’s T2DM OMOP query was applied to OMOP data and converted to create a consistent cohort using FHIR data. We ran a DQ assessment on each cohort, following the same approach used with the full study population.

Fitness of use was assessed using the Knowledge-Enhanced Electronic Profile Review (KEEPER) system, an OHDSI-maintained R package that extracts phenotype-related data elements along several key categories (Figure 2).⁴⁰ The categories included in this analysis were (1) clinical presentation, to extract symptoms present at cohort entry, (2) clinical plausibility, to extract relevant symptoms, diagnoses, treatment, or comorbidities prior to cohort entry, and exclude competing diagnoses after cohort entry, (3) diagnostic procedures, to extract relevant diagnostic procedures or laboratory tests, (4) treatment, to extract relevant medications prescribed on or after cohort entry, and (5) follow-up care and complications, to extract relevant complications after cohort entry. T2DM-relevant data elements were identified and extracted based on the Diabetes Mellitus Treatment Ontology,⁴¹ which contains curated structured definitions of diabetes treatment focused on data elements subsumed by diabetes symptom, diabetes drug, diabetes complication, and diabetes laboratory test nodes of the ontology to satisfy KEEPER-derived categories. To compare OMOP and FHIR, the proportion of participants in each standard that had at least one instance of each data element in both standards was calculated and a two-sample z-test was applied.

Figure 2.

Overview of KEEPER analysis. Shows which data elements and at which timepoints relative to cohort entry data were used for KEEPER analyses of T2DM cohort. This framework results in several prevalence analysis across domain concepts, as shown above.

Results

A total of 29 512 adult participants had at least one record in both OMOP and one record in FHIR.

Data comparison

Demographics

Comparison between FHIR and OMOP revealed nearly identical distributions for age, sex, and race (Table 1). Only ethnicity was significantly different between the two sources (P = .03), potentially due to data in each standard coming from different providers.

Table 1.

Sample demographics.

	OMOP		FHIR
	n	%	n	%	P-value
Age					.96
18-34	3260	11.1	3226	11.2
35-49	6629	22.5	6531	22.6
50-64	7525	25.5	7404	25.6
65-79	10 236	34.8	9959	34.5
80+	1809	6.1	1766	6.1
Excluded	53	0.2	625	2.2
Total:	29,459		28,887
Sex					.9
Female	18 342	62.7	18 057	62.6
Male	10 915	37.3	10 770	37.4
Excluded	255	0.9	34	0.1
Total:	29,257		28,827
Race					.37
American Indian or Alaska Native	88	0.4	84	0.3
Asian	735	3.0	704	2.9
Black or African American	1176	4.8	1133	4.6
Native Hawaiian or Other Pacific Islander	28	0.1	16	0.1
White	22 700	91.8	22 584	92.0
Excluded	4785	19.4	4991	20.4
Total:	24,727		24,521
Ethnicity					.03
Hispanic or Latino	1431	5.9	1377	5.4
Not Hispanic or Latino	22 793	94.1	23 943	94.2
Excluded	5288	21.8	4192	16.6
Total:	24,224		25,320

Bold indicates significant p-value

Healthcare encounters

Among all participants, only 1.5% had at least one FHIR Encounter, while 97.0% had OMOP visit_occurrence data. A total of 98 participants had at least one visit in either standard and were included in our encounter analysis.

Visit-days comparison showed greater visit-days in OMOP (median = 97) than FHIR (median = 10), while participants with visits in both standards had a median of 4 visit-days (Figure 3A).

Figure 3.

Encounter comparison. Comparison of records from the FHIR Encounter resource and the OMOP visit_occurence table, which both record patient interactions with a healthcare system. Only participants with existing FHIR encounters and OMOP visit_occurrences are included. (A) Comparison of the Visit-Days—defined as the number of days during which a participant starts a visit—per participant. Visit-Days only in OMOP, only in FHIR, and those that overlap between the two standards are shown. (B) Encounter and Visits by type. Of note, the Observation Encounter category is a specific encounter type that is only in FHIR (C) Durations (in days) of inpatient visits with both a start and end date populated. Durations are limited to less than 30 days.

Visit types comparison showed differing distributions between OMOP and FHIR (Figure 3B). While outpatient visits constituted the majority of visits in both OMOP (67.4%) and FHIR (58.3%), a considerable proportion of FHIR visits (27.4%) were categorized as Observation Encounters—defined as a period where a person is treated and monitored before determining their admission. This visit type has no equivalent in the OMOP vocabulary and thus was not observed in OMOP data. Differences in inpatient visits were also observed, with a larger proportion of OMOP visits (4.7%) classified as inpatient compared to FHIR (0.77%).

Among inpatient visits, the distribution of visit durations differed (Figure 3C). Within OMOP, the median visit duration was 3-days (mode = 1), while FHIR inpatient visit durations had a median of 4-days (mode = 2). For longer inpatient visits, the frequency of visits 30-days or longer was similar between both standards with the longest inpatient stay lasting 100-days in OMOP and 59-days in FHIR.

Codeable event volume

Most domains had comparable event counts per participant between OMOP and FHIR (Figure 4A). Based on median event counts, the most closely aligned domains were Measurement (OMOP = 846; FHIR = 832), Drug (OMOP = 92; FHIR = 90), and Observation (OMOP = 65; FHIR = 100), while median event counts within OMOP were considerably higher in the Condition (OMOP = 258; FHIR = 11) and Procedure (OMOP = 72; FHIR = 4) domains.

Figure 4.

Comparison of codeable event volume between OMOP and FHIR. (A) Comparison of OMOP and FHIR volume of clinical events by domain per person. In OMOP, we define an event as a table row with a single domain-mappable code field. In FHIR, we define an event as the unique occurrence of a code in a resource, such that one resource can contain multiple events. We borrow OMOP’s concept mapping conditions to classify FHIR codes into domains. (B) The number of providers submitting clinical events per person.

Comparison of providers per participant revealed that while the vast majority of data in both standards came from a single provider, considerably more participants in FHIR than OMOP had data from multiple providers (Figure 4B). In OMOP, only 20 participants held data from multiple providers, with a maximum of 2 providers while 6300 participants had FHIR data from multiple providers, with a maximum of 13 providers.

Timeline comparison

Comparison of record dates showed a considerable number of records were only available within either OMOP or FHIR (Figure 5). Among participants with at least 10 records (n = 17 123) in both FHIR and OMOP, only 15.9% of records on average fell on the same date in both standards. On average, 65.3% of records were unique to OMOP, while 18.4% were unique to FHIR.

Figure 5.

Person Timeline of FHIR and OMOP records by date. A visualization of FHIR and OMOP Drug, Visit, Condition and Procedure event start-dates for the top 15 participants with the most events in both standards. The participants were chosen by calculating the number of records in each standard per participant, selecting the minimum of the two, sorting in descending order, and selecting the top 15. Recorded event dates uniquely found in FHIR, uniquely found in OMOP, and found in both show each standard may capture different events in a participant’s EHR.

Data quality

Of the 41 metrics tested, 25 (61%) showed a significant difference in DQ performance between OMOP and FHIR (see Supplementary Materials). The greatest difference was observed among data completeness metrics, with 11 (61%) of the 18 metrics showing a significantly higher (P < .001) proportion of data completeness in OMOP (63-100%) compared to FHIR (0-100%). Within FHIR, data completeness was lowest among resources containing non-null references to Encounter resources. A similar pattern emerged with references to Practitioners and Locations, with only 18% of procedures in FHIR containing a practitioner reference, compared to 63% in OMOP.

Of the 17 conformance metrics run, 9 (53%) revealed a significant difference between the two sources (P < .001), with FHIR marginally outperforming OMOP (99-100%). Difference in performance was most pronounced for the procedure_reference_practitioner_resolved and medication_request_reference_encounter_resolved metrics which assess referential integrity of resources, suggesting that when a resource includes a reference, FHIR will almost always be able to resolve it.

Of the five plausibility metrics run, three (60%) showed a significant difference between OMOP and FHIR (P < .001). The most pronounced difference was observed with the plausibleUnitConceptIds metric, which checks for compliant units for measurements and was satisfied for 96% of eligible FHIR records, but only 71% of eligible OMOP records. Within OMOP data, the best performing metrics were plausibleValueLow (100%), and plausibleTemporalAfter (100%). When assessing plausibleValueHigh, FHIR (100%) and OMOP (99%) were comparable.

Phenotype analysis

Our phenotype analysis included 3208 participants with T2DM identified in OMOP, 1979 identified in FHIR, and 1179 overlapping between the two. OMOP data came from an average of 1 provider per participant with a maximum of 2 providers, while FHIR data came from an average of 1 provider per participant with a maximum of 3 providers. Our DQ assessment of T2DM cohort data showed similar results to the full study population (see Supplementary Materials).

Results from KEEPER revealed several statistically significant differences between the two data sources (Figure 6). When looking at clinical presentation, less than 5% of participants in both cohorts reported symptoms upon cohort entry. The symptoms reported, however, differed significantly between the two standards, with chest pain (OMOP = 3.1%, FHIR = .55%, P < .001) and dyspnea (OMOP = 2.2%, FHIR = .91%, P < .001) most frequently reported in OMOP, while overweight (OMOP = 0%, FHIR = .45%, P < .001) and low testosterone (OMOP= < .01%, FHIR = .45%, P < .001) were most frequent in FHIR.

Figure 6.

Comparison relevant Type-II diabetes mellitus variables between OMOP and FHIR. *Indicates significant difference between FHIR and OMOP.

Results from the clinical plausibility category showed a greater proportion of medications prescribed prior to FHIR cohort entry relative to OMOP. This was particularly the case for the drug metformin (OMOP = 19.4%; FHIR = 54.8%, P < .001). Conversely, symptoms reported prior to cohort entry were greater within the OMOP cohort relative to FHIR, with chest pain (OMOP = 25.2%, FHIR = 3.1%, P < .001) and dyspnea (OMOP = 20.1%, FHIR = 2.7%, P < .001) being the most prevalent. Similarly, the proportion of comorbidities reported prior to cohort entry was greater in OMOP with essential hypertension (OMOP = 45.0%, FHIR = 24.3%, P < .001) and obesity (OMOP = 30.7%, FHIR = 11.6%, P < .001) as the top reported comorbidities in OMOP, while hypertensive disorder (OMOP = 4.8%, FHIR = 13.4%, P < .001) was most frequent in FHIR. The proportion of competing diagnoses following cohort entry were also higher in OMOP, with prediabetes (OMOP = 14.9%, FHIR = 4.2%, P < .001) and impaired glucose tolerance (OMOP = 3.5%, FHIR = 0.4%, P < .001) both being higher in OMOP relative to FHIR.

Discussion

The current study characterizes real-world data in the OMOP and FHIR standards using a shared pool of participants from the AoURP. Findings revealed some overlap between data in the two standards, but with several differences. One key difference is the data volume across domains, with OMOP having significantly greater data density in the Condition and Procedure domains relative to FHIR. Similarly, our T2DM analysis showed higher density of symptoms, complications and comorbidities data in OMOP, while FHIR contained more medications data. Though OMOP showed a higher level of data volume, FHIR was able to capture data across multiple health systems and often included additional timeline segments not present in OMOP, suggesting FHIR can supplement OMOP with data from visits occurring beyond the single institution OMOP is limited to.

The finding of greater data volume in OMOP in nearly all evaluated domains was somewhat surprising as more participants had multiple data sources in FHIR compared to OMOP. This discrepancy is potentially due to the ETL process required of OMOP and the iterative AoURP data submission process. Through the ETL process, engineers can map concepts from anywhere in the EHR to the OMOP standard vocabulary, giving them flexibility to adjust concept mappings based on institutional coding practices.²⁰ Conversely, FHIR is limited to standard mappings used in its default bulk export functionality. Since these mappings are determined by the EHR vendor, clinical data documented in an unexpected area of the EHR are often excluded from FHIR exports, potentially explaining FHIR’s lower data density. For this reason, anyone considering how to adopt the data models should consider the level of granularity needed as the flexibility of OMOP’s ETL process enables better capture of granular data elements. For example, blood transfusion data may not appear in FHIR payloads, but mappings already exist to populate OMOP using Information Standard for Blood and Transplant codes.⁴²

OMOP also displayed greater data coverage within our T2DM cohorts. Notably, fewer individuals were identified in FHIR compared to OMOP. This is potentially due to higher prevalence of diabetes-linked symptoms, complications, comorbidities and measurements in OMOP which likely resulted in more qualifying condition events in OMOP relative to FHIR. Interestingly, the prevalence of medication data was higher in FHIR, particularly for common T2DM drugs like metformin and insulin lispro. This is potentially due to FHIR data coming from more data sources relative to OMOP, which may have enabled better capture of medication data in FHIR.

Our DQ assessment showed OMOP outperformed FHIR on several completeness metrics. FHIR was especially poor at defining encounters linked to resources. This is likely due to poor capture of encounters within FHIR, as FHIR slightly outperformed OMOP when evaluating the resolution of the few encounters that were available in FHIR. In these cases, FHIR seems less complete, but guarantees stronger conformance to its data model. FHIR’s strong conformance was further demonstrated by its ability to ensure unique primary keys, an area where OMOP fell short. This is likely due to FHIR’s application of strict API-enabled validation checks at time of server ingestion—often blocking data loads—which likely improves conformance relative to OMOP’s non-blocking and ad-hoc methods of checking DQ. Overall, these results suggest a tradeoff between data yield and data conformance, marking a critical difference between the two standards.

When considering data conformance, it is also worth discussing the enforcement of standard concepts. While mapping of non-standard concepts to OHDSI’s standardized vocabulary is foundational to OMOP, FHIR has much looser standardization requirements. Although it defines internal and external terminologies, such as SNOMED-CT, ICD10, and RxNorm, FHIR rarely enforces a strict adherence. This exposes a critical weakness in an otherwise very conformant data model. The AoURP circumvents this issue by using a FHIR data aggregator, such as Care Evolution, to perform probabilistic and deterministic mappings of text-based codes to standard terminologies.⁴³ This process has resulted in FHIR codes having an even higher conformance to its standard vocabularies. Importantly, our analyses were only possible because this enhancement was performed, underscoring the need for such enhancements to become commonplace if FHIR is to be used for data analytics.

Perhaps the most notable difference between OMOP and FHIR was encounter reporting, with only 1.5% of participants having a FHIR encounter record, compared to 97% in OMOP. This likely reflects current USCDI requirements, which does not explicitly require encounter information.⁴⁴ While later versions include an encounter requirement, their implementation will not be mandated until 2026.⁴⁵ As a consequence, it seems that minimal compliance, which excludes encounters, is the predominant practice. Additionally, we observed a mix of EHR vendors that have encounters, with most NextGen vendors having encounter profiles, but this was inconsistent across Epic payloads. Despite this, stricter USCDI regulatory requirements are clearly needed to improve encounter reporting.

Interestingly, when FHIR encounters were reported, they had little overlap with reported OMOP encounters. Notably, the majority of visit-days occur strictly within OMOP, then strictly in FHIR, and lastly in an overlap between the two. This implies a low level of “redundancy” between the datasets. This was evident in our timeline comparison which showed that most records were either only available in OMOP or only available within FHIR. Thus, while OMOP covered more days overall, FHIR was able to capture visits not available within OMOP. These findings suggest that FHIR could be used to supplement OMOP by providing data from visits occurring beyond OMOP’s single-site dataset.

Limitations and future directions

The present study is not without limitations. First, our data was limited to a subset of participants from a single research consortium. As such, our results are influenced by AoURP’s data submission processes, which were developed specifically for the AoURP following OMOP best practices and using third party FHIR preparation methodology. Additionally, only data submitted between January 2020 and March 2024 was used, meaning updates in how data are captured are not reflected within our analysis. Additional work is therefore needed to compare OMOP and FHIR data across research programs with differing data preparation and submission requirements.

Second, our comparison of OMOP and FHIR only included a small subset of available data elements. Given that both standards are capable of representing thousands of clinical data elements, future analysis should aim to develop more robust OMOP to FHIR mappings to enable a more extensive comparison. Furthermore, our domain selection focused on core OMOP research tables, which may inherently align more closely with OMOP’s design and intended use. As a result, capabilities of FHIR outside these domains were not evaluated and may be underrepresented in this comparison, potentially introducing a bias. Future work may focus more on data structures in FHIR that are not readily available in a typical OMOP repository. Third, our encounter comparison focused on visit-specific attributes, so a logical next step is to comprehensively assess encounters (eg, incorporating codeable events) across OMOP and FHIR. Finally, our cohort analysis only evaluated a single phenotype algorithm that was developed and validated based on OMOP data. Though a FHIR-based algorithm was developed for this analysis, more work is needed to develop and validate phenotype algorithms using FHIR to enable greater use of FHIR for research purposes. Additionally, future work should look to compare the fitness for use of OMOP and FHIR datasets on multiple phenotype algorithms from PheKB³⁹ or other knowledge sources such as the OHDSI Phenotype Library.⁴⁶

Conclusion

Overall, this comparison of OMOP and FHIR suggests both standards can support data analytics and real-world evidence generation. Our results, however, suggest OMOP offers certain advantages for clinical research due to its higher levels of data completeness, codeable event volume, and availability of encounter data—though we expect this gap to lessen as USCDI mandates become stricter. Notably, FHIR is able to capture clinical encounters occurring beyond a single health system and contains participant timeline segments not present in OMOP, suggesting that using OMOP and FHIR in tandem could offer a fuller picture of each participant’s healthcare journey.

Author contributions

Jason S. Patterson (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Validation, Visualization, Writing—original draft, Writing—review & editing), Elise Minto (Conceptualization, Data curation, Formal analysis, Methodology, Validation, Visualization, Writing—original draft, Writing—review & editing), Maura Beaton (Conceptualization, Project administration, Supervision, Visualization, Writing—original draft, Writing—review & editing), Adit Anand (Writing—review & editing), Mark Velez (Data curation, Resources), Paul A. Harris (Resources, Supervision, Writing—review & editing), George Hripcsak (Supervision, Writing—review & editing), and Karthik Natarajan (Funding acquisition, Supervision, Writing—review & editing)

Supplementary material

Supplementary material is available at Journal of the American Medical Informatics Association online.

Funding

This work was supported by funding from the National Institutes of Health, Office of the Director grant number 5 U2C OD023196.

Conflicts of interest

None declared.

Data availability

The data underlying this article contains personal health information and cannot be shared publicly. A de-identified dataset can be shared on reasonable request to the corresponding author.