Bridging large-scale FHIR to AI in production
summary
Describes an "assembly line" for turning large-scale FHIR into production AI: canonical ML pipelines (example: Vertex Pipelines) that read FHIR/ DICOM, produce and persist model metadata store entries and metrics, and push models to a prediction service with an ML sidecar for inference. Demonstrates an end‑to‑end real‑time risk‑stratification system, explains storing training cohort references (not raw metadata) back to FHIR, and mandates per‑patient confidence scores, cohort calibration, and continuous monitoring.
AI-written summary · about
chapters (11)
- 0:00Introduction and motivation
- 2:11AI interoperability challenges in FHIR
- 4:23Standardization in three areas
- 10:21Transforming EHR data for ML
- 12:51Handling image data in AI workflows
- 14:17Model architecture and metadata
- 15:45Metrics and risk stratification
- 19:15Live inference and confidence scores
- 21:46MLOps and deployment pipelines
- 25:19Open challenges and future work
- 26:01Q&A session
related talks
39mAI, FHIR & the future of healthcare data
30mFrom Code to Clinic — Building an Internal Developer Platform for Clinical AI
48mAnalytics and AI on FHIR
11mDesigning Governed AI Pipelines for Healthcare Impact
30mFHIR in the W.H.O.L.E.: Graphing Clinical Data for Smarter AI Systems
19mAnalyzing FHIR Data Using ML and Analytical Tools
54mArtificial Intelligence seen from the software development lifecycle perspective
36mResponsible AI