How to Implement Named Entity Recognition, Relation Extraction etc. for a Custom FHIR Profile Using Azure?

Hi Jack K,

Thanks for reaching out to Microsoft Q&A.

TLDR: At present, there is no direct support in Azure Text Analytics for Health for custom FHIR profiles, so implementing a mapping and post-processing strategy is the best approach to handle the extensions specific to your region.

====

To implement Named Entity Recognition (NER) and relation extraction for a custom FHIR profile using Azure Text Analytics for Health, here's a strategy you can follow

1. Understand the Base Profile: Azure Text Analytics for Health uses standard FHIR profiles as a baseline for entity extraction. However, custom FHIR profiles are not directly supported for entity extraction out of the box.
2. Extract Standard FHIR Entities: Use Azure's Text Analytics for Health to extract standard FHIR entities, such as medical conditions, medications, treatments, and more. These entities are based on international healthcare standards like ICD-10, SNOMED, and RxNorm, aligned with FHIR. You can leverage this feature to capture the basic entities relevant to your custom profile.
3. Post-Processing for Custom FHIR Entities: Since the Azure service does not natively support country-specific FHIR extensions, you will need to perform post-processing to map extracted entities from the standard profile to your custom FHIR profile. You can use the following techniques:
   • Custom Mapping Logic: Create custom logic that takes the extracted standard FHIR entities and maps them to your country's FHIR extensions. This can be done by defining a mapping layer within your application that transforms standard entities to the custom entities in your profile.
   • Custom Relation Extraction: If your custom FHIR profile includes additional relationships between entities, you may need to use Azure's Text Analytics for Health's relation extraction capabilities and supplement them with custom rules or models specific to your country's profile. For example, develop a custom machine learning model or rules-based system to recognize these specific relationships in the text.
4. Custom NER with Azure Machine Learning (Optional): If your custom FHIR profile extends significantly from the standard FHIR entities, you may need to train your own NER model using Azure Machine Learning. This model would be tailored to extract entities defined in your country's custom FHIR profile. Azure ML allows for custom model deployment, and you can integrate these models with the Text Analytics for Health API to ensure both standard and custom entities are recognized.
5. Custom FHIR Resource Generation: After extraction, based on the output entities and relationships from Azure Text Analytics, you can build custom FHIR resources by transforming them into the specific format required by your country's custom profile. This could involve enriching the data with additional fields, modifying existing fields, or even restructuring relationships between resources to match the custom profile.
6. Validation of FHIR Resources: After generating custom FHIR resources, validate them against your country's FHIR specification to ensure they conform to the extended profile. You might need to use a local FHIR validation tool or a custom validation service that supports your country-specific FHIR rules.

Please 'Upvote'(Thumbs-up) and 'Accept' as an answer if the reply was helpful. This will benefit other community members who face the same issue.