Grad Seminar: Encoding FHIR Medical Data for Transformer Models
Grad Seminar: Encoding FHIR medical data for transformer models - University of Waterloo
Summary
A student from the University of Waterloo presented a method utilizing the open-source HL7 FHIR standard to generate token IDs from electronic health record data for machine learning. This method was validated on the MIMIC-IV FHIR dataset for tasks such as length-of-stay and mortality prediction.
Details
This seminar addresses challenges in developing ML solutions using medical data, citing issues like reliance on bespoke database schemas and closed-source datasets due to privacy concerns. It highlights the advantage of the open-source FHIR standard for interoperability across various platforms (EHR storage, payment/insurance, ML analytics). The core technical contribution is a novel data encoding method: tokenizing JSON returned from an FHIR server query into a sequence of token IDs based on the underlying FHIR data structure. These token IDs are designed to train transformer language models on medical records. Performance was validated using the open-source MIMIC-IV FHIR dataset for specific tasks, including length-of-stay prediction and mortality prediction. Furthermore, the presentation explores potential solutions for limitations such as handling large sequence lengths and the discrete tokenization of continuous numeric data. This work provides a concrete technical framework for building ML training pipelines based on the FHIR standard, which is highly relevant to Japanese healthcare IT where standardization and interoperability are critical for AI development.
Original content copyright by respective publishers