TL;DR
This paper tackles the challenge of making rich, real-world clinical data interoperable by converting the extensive MIMIC-IV database into the FHIR standard. They've not only created this FHIR-formatted dataset but also released a demo version and an implementation guide, paving the way for broader research and development in healthcare informatics.
What Problem Does This Paper Address?
The healthcare industry is awash in digital data, but this data is often siloed within proprietary systems, making it incredibly difficult to share and analyze across different institutions or applications. This lack of interoperability hinders research, slows down the development of new healthcare tools, and limits our ability to gain deeper insights from patient information. While standards like HL7 FHIR (Fast Healthcare Interoperability Resources) are designed to solve this, the real-world application of FHIR to large, complex clinical datasets has been limited. Specifically, there was a need for a de-identified, real-world critical care dataset that was not only available in FHIR but also accompanied by clear guidance on its structure and usage.
Methods
The authors employed a four-stage process to convert the MIMIC-IV database into FHIR:
- Model: They first defined FHIR profiles and terminology mappings based on the base FHIR R4 resources, heavily influenced by the US Core standard. This involved creating constraints and extensions to accurately represent the MIMIC-IV data.
- Generate: The relational data from MIMIC-IV was then transformed into FHIR resources according to the defined profiles. This was achieved using custom SQL scripts and mapping MIMIC-IV data types to FHIR JSON data types.
- Validate: To ensure accuracy and conformance, the generated FHIR resources were validated against the defined FHIR profiles using an open-source FHIR server (HAPI-FHIR). This step was crucial for identifying and correcting any inconsistencies or errors.
- Publish: Finally, the validated FHIR resources were published as newline-delimited JSON (NDJSON) files. An implementation guide (IG) was also created, containing all the FHIR profiles, terminology, and documentation, making it easier for others to understand and use the data.
Datasets
The primary dataset used in this work is the MIMIC-IV database, a de-identified, real-world critical care dataset containing data from over 315,000 ICU stays.
The authors created two versions of the FHIR-formatted MIMIC-IV dataset:
- MIMIC-IV on FHIR (full dataset): Contains approximately 5,840,000 resources for 315,000 patients.
- MIMIC-IV on FHIR (demo dataset): A subset of 100 patients with over 915,000 resources, made openly available to facilitate broad reuse and rapid prototyping.
This work builds upon previous efforts in creating synthetic FHIR datasets (like Synthea) but distinguishes itself by using real-world, de-identified patient data.
Key Insights and Ideas
- Real-world data in FHIR is a game-changer: While synthetic data is useful, having a large, de-identified, real-world clinical dataset in FHIR unlocks new possibilities for research and development that synthetic data cannot replicate.
- Terminology mapping is a significant challenge: Accurately mapping the diverse terminologies found in a real-world EHR to FHIR standards requires careful consideration and can be a complex undertaking.
- FHIR's verbosity comes at a storage cost: Converting relational data to FHIR resources significantly increases storage requirements, which can be a barrier for some use cases.
- Validation is critical but computationally expensive: Thorough validation of FHIR resources against profiles is essential for ensuring data quality but can be a time-consuming process, especially for large datasets.
- Open-source tools and community collaboration are key: The use of open-source tools (like FSH, HAPI-FHIR, and the FHIR IG Publisher) and the release of an open-source repository for the conversion process are vital for fostering community adoption and further development.
Key Results
- 25 FHIR profiles were derived from the base FHIR R4 resources.
- 34 CodeSystems and 35 ValueSets were generated to represent MIMIC-IV's terminology.
- The demo dataset contains 100 patients and over 915,000 resources.
- The full dataset comprises 315,000 patients and approximately 5,840,000 resources.
- Storing MIMIC-IV data in FHIR format within a PostgreSQL database resulted in a 7.5x increase in storage compared to the original relational structure. This increased to a 49x increase when the FHIR data was inserted into a FHIR server.
- The demo resources were generated in PostgreSQL in 50 seconds, with validation taking 112 minutes.
- The MIMIC IG includes 24 profiles, 34 CodeSystems, 37 ValueSets, 2 extensions, and extensive documentation.
Limitations and Future Work
- US Core mapping: The authors did not map to the US Core standard directly, as it would require extensive terminology mapping. They believe that mapping the generated resources to US Core using FHIR tools will result in a more broadly usable approach in the future.
- Data curation: The dataset was not curated, meaning it retains biases present in routine clinical data collection. Researchers need to be aware of this context.
- Imputed values: For some resources, required elements were missing and imputed with a value of "unknown."
- Other MIMIC-IV modalities: The work did not include mapping other MIMIC-IV data modalities like MIMIC-CXR, the MIMIC Waveform Database, or MIMIC-IV-Note. Future work could extend to these areas.
- Storage optimization: While FHIR offers interoperability benefits, the increased storage requirements are a significant consideration. Future research could explore more efficient storage formats for FHIR data.
- Validation efficiency: The validation process for large datasets is computationally expensive. Exploring parallelization approaches or leveraging cloud-based FHIR servers could improve efficiency.
Summary updated