HCC Retrospective Coding

Coding & Risk AdjustmentAsync jobs required

Retrospective HCC coding with MEAT evidence, V24/V28 mapping, and governance checks.

HCCRisk adjustmentRAFMEATICD-10

About

Codes risk-adjustable conditions from historical charts for retrospective review. A two-pass pipeline first extracts every documented diagnosis, then shortlists and validates each candidate against MEAT criteria (Monitor, Evaluate, Assess, Treat), maps it to CMS-HCC V24 and V28 categories, and runs a governance checklist.

Every returned code carries its evidence trail: supporting sentences with document name, page number, and normalized bounding boxes, plus per-category MEAT evidence — so reviewers can click from a code straight to the exact line in the chart that supports it.

How it works

  1. 1Document upload → OCR (skipped for the /text variant)
  2. 2Pass 1 — diagnosis extraction across all dates of service
  3. 3Pass 2 — shortlisting, MEAT validation, HCC V24/V28 enrichment, governance rules
  4. 4Optional reference_codes input steers specificity matching against codes you already hold

Intended use

  • •Retrospective chart review and RAF (risk adjustment factor) capture programs
  • •Routed to when users ask for "HCC retrospective", "RAF", or "risk adjustment"
  • •Coder workbenches with split-panel evidence UIs (code list left, highlighted chart right)

Key outputs

  • ▸result.ICD_descriptions[] — codes grouped by date of service with page ranges
  • ▸Per code: hcc_v24 / hcc_v28 category, is_hcc_relevant, plain-language reason
  • ▸meat_evidence — monitor/evaluate/assess/treat citations, each with page + bbox
  • ▸governance_flags & governance_checklist — F2F source, provider credential validity, PMH-only and lab-inference flags

Model comparison

F1 on Gwen's healthcare benchmark for this task — the Gwen pipeline vs the prompt-optimized model alone, with the uplift the pipeline adds, per model.

#ModelGwen pipelineModel onlyUplift
1
GPT-5.5Best
0.961
0.919+0.042
2
Gemini 3.5 Flash
0.953
0.857+0.096
3
Claude Opus 4.8
0.947
0.945+0.002

Endpoints

Try each endpoint with your signed-in session — usage counts toward your monthly budget.

Use synthetic data only. Do not submit real patient records or PHI when testing endpoints.

Limitations & caveats

  • –Generic "medical coding" requests belong to ICD Coding, not HCC — this service is risk-adjustment-specific
  • –Retrospective scope: codes every MEAT-supported documented condition; for point-of-care coding of actively addressed conditions use HCC Prospective
  • –Runs 1–3 minutes; synchronous calls will time out — the async /jobs flow is mandatory
  • –Bounding boxes are normalized 0–1 relative to the page (top-left origin)