Design a Medical Case ICD-10 Suggestion System
Generate versioned, reviewable diagnosis-code suggestions from clinical cases without bypassing human coding decisions.
Problem statement
For this exercise, assume a health system submits a finalized, versioned clinical case containing structured encounter facts and authorized note text.
The service creates a ranked ICD-10 suggestion set with calibrated confidence and supporting evidence references, then a certified coder accepts, rejects, or replaces suggestions before any billing export. A case version, terminology release, feature pipeline, and model version must make every result reproducible.
Submissions and reviews are idempotent, protected health information stays inside an audited tenant boundary, and model output can never silently overwrite a human decision.
Support asynchronous inference, status reads, review, quality sampling, safe model rollout, reprocessing into a separate suggestion version, and reconciliation. Clinical diagnosis, autonomous billing, model training, and claims submission are outside scope.
What to build
- Submit one authorized clinical-case version and obtain a retry-safe inference job.
- Return ranked ICD-10 suggestions with confidence, evidence references, and exact terminology and model versions.
- Let a certified coder accept, reject, or replace suggestions while preserving an immutable review audit.
- Roll out or roll back model versions and reprocess cases without rewriting prior suggestions or decisions.
- Measure latency and coding quality without exposing protected health information outside its tenant boundary.
At scale
- Assume 40 million case versions per year, 1,500 submissions per second at peak, and note payloads up to 2 MB.
- A normal suggestion job should finish within 15 seconds p95; status reads and review writes within 300 ms p95.
- Suggestions, reviews, model lineage, and access audit are retained for seven years.
- A retry, worker crash, or model rollout cannot create two current authoritative decisions for one case version or mix model artifacts within one suggestion set; superseded reviews remain immutable history.