Granular Segmentation Benchmark on Synthetic FHIR Cohort

Benchmarking consent thresholds across 11,519 synthetic patients for throughput, accuracy, and cost.

Why I Built This

Privacy-preserving architectures are easy to pitch, but hard to operationalize without hard numbers. I built this benchmark to measure what happens when consent-aware segmentation runs at cohort scale, including latency, throughput, and cost behavior. This gave us concrete evidence for design tradeoffs instead of assumptions.

What I Evaluated

  • 11,519 synthetic patient records generated with Synthea.
  • 197.6 million FHIR resources processed across 11 confidentiality thresholds.
  • Physician review of sensitivity classifications and segmentation accuracy.
  • Throughput, scalability, and estimated AWS Lambda runtime cost.

Outcome

SHARES processed about 18,900 FHIR resources per second on a standard laptop. Across the full evaluation, 8.64% of resources were flagged as sensitive, and a complete segmentation run cost less than one cent.