Generating FHIR Datasets from Clinical Notes: A Synthetic Approach
Synthetic Data for a Notes-to-FHIR Model | by Mostafa Davoodi | Jul, 2026 | Medium
Summary
The author details a method for synthetically creating paired datasets of clinical notes and structured healthcare data in the FHIR format. This approach allows for the construction of necessary training data for fine-tuning small, private language models without relying on sensitive real patient records.
Details
This article provides a technical walkthrough on developing a small language model that takes a provider's clinical note as input and outputs the corresponding structured JSON in FHIR format. A major hurdle in traditional medical AI is obtaining paired training data because real notes are protected, making manual labeling slow and difficult. To overcome this, the author employs a synthetic data generation approach. The process involves first generating fully simulated patient records using Synthea (which produces valid FHIR). From these full records, only necessary fields (name, birth date, gender, etc.) are extracted to create a
Original content copyright by respective publishers