HCC risk adjustment automation for ACOs and Medicare Advantage.
Pre-visit clinical NLP that surfaces unaddressed HCC conditions from EHR notes, in time for the clinician to confirm them during the encounter. Built on Azure OpenAI, Microsoft Fabric, and Claude API. Designed for ACO REACH, MSSP, Medicare Advantage, and commercial risk-bearing contracts.
Why HCC accuracy is now a board-level conversation
CMS V28 fully phases in by performance year 2026, removing more than 2,000 ICD-10 codes from the HCC model and lowering coefficient values across many remaining conditions. Medicare Advantage plans and risk-bearing ACOs that have been carrying mature retrospective coding programs are finding that the V24 playbook simply does not produce the same financial result under V28. Layer in tighter RADV scrutiny, increased payer audit activity, and the natural drift between prior-year documented conditions and current-year clinical reality, and risk adjustment accuracy is now a board-level conversation rather than a coding department conversation.
The organizations that perform under V28 are the ones that surface unaddressed conditions to the clinician at the point of care, with structured evidence, every visit. That requires production-grade clinical NLP. Not a chart review service. Not a quarterly sweep. A workflow.
What HCC risk adjustment automation actually involves
The pipeline is the easy part. The hard part is clinical workflow integration, accuracy validation, and the operating model that turns NLP suggestions into confirmed conditions.
Note ingestion and preprocessing
Pull progress notes, consult notes, and discharge summaries from Epic, Cerner, or Meditech via FHIR DocumentReference and DiagnosticReport. Normalize section headers, redact non-clinical content, and tokenize for downstream models. Track lineage to the source note and source author for audit defensibility.
HCC extraction with confidence scoring
Two-stage extraction. First pass: candidate condition mentions with ICD-10 codes. Second pass: HCC mapping under V24 and V28 with a confidence score. Models tuned to your EHR note style on a labeled gold-standard set. Threshold tuning per HCC category based on your tolerance for false positives in clinician workflow.
Pre-visit clinician surfacing
Surfaced to the clinician 24 to 48 hours before the encounter, embedded in their daily workflow (in-EHR worklist, MyChart-attached note, or standalone surfacing app, depending on EHR posture). Each suggestion shows the source-note evidence, the confidence score, and a one-click confirm or reject affordance.
Audit trail and validation harness
Every suggestion logged with the source note evidence, the confidence score, the clinician decision, and the timestamp. A continuous validation harness scores model accuracy on a held-out set monthly. The audit trail is structured the way RADV and commercial payer auditors want to see it.
How we deliver
One senior team end to end. Clinical informatics in the room from week one. Production-grade evaluation from before the first model call.
- 01
Discovery and ROI modeling (2 to 3 weeks)
Inventory your EHR estate, attribution model, current chart-review program, and prior-year coding baseline. Sample notes to estimate addressable HCC opportunity under V24 and V28. Build a defensible per-attributed-life ROI estimate. Output: scoped statement of work and an ROI model your CFO can sign off on.
- 02
Gold-standard build and model selection (3 to 5 weeks)
Clinical informatics builds a labeled gold-standard set on real notes. Benchmark Azure OpenAI, Claude API, and open-source candidates for precision and recall on your note style. Model selection is data-driven, not vendor-driven. Document the architecture and data residency model.
- 03
Pipeline build and clinician surfacing (8 to 12 weeks)
FHIR-based note ingestion, two-stage extraction, confidence scoring, clinician surfacing UI integrated with the existing workflow. Embed audit trail and dead-letter handling from day one. Soft launch to a single specialty or pod.
- 04
Clinical-accuracy validation and full rollout (4 to 6 weeks)
Monthly precision and recall measurement on the held-out set. Threshold tuning per HCC category. Clinician feedback loop. Expand from pilot pod to full attributed population once accuracy and adoption hold.
- 05
Operations, RADV-readiness, and ongoing support
Quarterly model retraining cadence. Audit-trail review. CMS V28 transition runbook. Optional managed support if your team is small. Annual ROI reconciliation against the original model.
What you get
- Production HCC NLP pipeline live against your primary EHR (Epic, Cerner, or Meditech)
- Labeled gold-standard set on your real notes, owned by you
- Two-stage extraction with V24 and V28 HCC mapping and confidence scoring
- Pre-visit surfacing integrated with clinician workflow (in-EHR or attached app)
- Per-HCC-category accuracy thresholds tuned to your false-positive tolerance
- Defensible audit trail with source evidence, confidence, and decision lineage
- Continuous validation harness with monthly precision and recall reporting
- Modeled per-attributed-life ROI baseline plus annual reconciliation
- CMS V28 transition runbook and coefficient-impact analysis
- Optional managed support and quarterly retraining cadence
When to engage us
Your V28 forecast does not hold
If your Medicare Advantage or ACO REACH revenue forecast under V28 is missing the V24 baseline, retrospective coding alone will not close the gap. NLP at point of care is now the lever that has scale.
Your retrospective program has plateaued
If your chart-review program is mature but year-over-year capture is flat, you have hit the ceiling of retrospective work. Pre-visit surfacing pulls a different curve.
You are entering ACO REACH or downside-risk
Risk-bearing entry without HCC NLP infrastructure is a structural disadvantage. Build the workflow before performance year one, not during it.
You are facing a RADV or commercial audit
Strong audit trails are easier to build into a new pipeline than to retrofit. If audit defensibility is becoming a board concern, the pipeline structure matters as much as the accuracy.
Pitfalls we see in HCC NLP projects gone sideways
- Treating model precision as a single number. Per-HCC-category precision varies by 15 to 25 points. Threshold tuning per category is the difference between clinician adoption and clinician backlash.
- Surfacing without source-note evidence. Clinicians will not confirm a suggestion they cannot trace back to a note. This breaks workflow and audit trail in one move.
- Ignoring the V24 to V28 transition. A model trained against V24 mappings will silently miss conditions whose codes shifted under V28. Build with both maps from day one.
- Skipping the gold-standard set. Vendor accuracy claims are not your accuracy. Without your own labeled set you cannot tune, you cannot validate, and you cannot defend.
- Bolting NLP onto a chart-review team and calling it integrated. If the suggestion does not reach the clinician before the visit, you have a more expensive retrospective tool. The workflow integration is the work.
Related reading
HCC risk adjustment NLP, the practitioner guide
Architecture, model selection, accuracy validation, and clinician workflow integration for production HCC NLP.
FHIR integration consulting
FHIR-native note ingestion is the foundation under HCC NLP. The pipeline that powers risk adjustment also powers everything else.
What the CMS ACCESS Model actually requires
FHIR APIs, PROMs, and outcome attainment requirements broken down for ACOs preparing for January 2026.
Frequently asked questions
What is HCC risk adjustment automation, and what does it actually replace?
HCC risk adjustment automation uses clinical NLP to surface previously documented but currently unaddressed Hierarchical Condition Categories from EHR notes, before the patient encounter. It replaces retrospective chart review, where coders comb through prior-year notes looking for missed conditions, and surfaces the same conditions to the clinician in time to confirm them during a live encounter. The economic effect is the same. The clinical effect is significantly better, because the condition gets addressed in care instead of in coding.
How accurate is HCC NLP today, and how do you handle clinical accuracy review?
Modern clinical NLP on production data routinely hits 85 to 92 percent precision on common HCC categories when properly tuned to the EHR's note style. Our deployments include a clinical-accuracy validation harness against a labeled gold-standard set, a reviewing-clinician workflow for low-confidence surfacing, and a feedback loop that retrains thresholds on rejected suggestions. The threshold for surfacing to a clinician is calibrated to your tolerance for false positives in workflow.
What's the typical ROI on HCC risk adjustment automation?
On Medicare Advantage and ACO REACH populations, we typically see $300 to $900 per attributed life per year in improved coding accuracy net of program costs, with the wider range driven by panel acuity, prior coding maturity, and how aggressive the existing chart-review program already was. On MSSP populations, the math is more about benchmark accuracy than payment, but the directional effect on shared savings is meaningful. We model expected ROI on your specific panel before kickoff.
Will my coders still have a job after this is in production?
Yes. The work shifts from retrospective review to oversight, validation, and edge cases. Coders become higher-leverage. They review NLP-surfaced suggestions, handle the conditions the model is uncertain about, and own the audit trail that auditors and CMS care about. Our deployments are designed around clinician and coder workflow integration, not replacement.
How does this hold up under a CMS RADV audit or commercial payer audit?
Auditability is built into the deployment. Every NLP suggestion ships with the source note evidence, a confidence score, and a clinician confirmation record. The audit trail is structured the way auditors want it. We have specifically designed deployments to be defensible under RADV-style review, and we will show you the audit-trail design as part of discovery.
Do you build on Azure OpenAI, Claude API, or open-source models?
All three, depending on the constraint. Azure OpenAI for HIPAA-aligned managed deployments inside Microsoft tenants. Claude API for clinical reasoning quality on harder note styles. Open-source models (typically Llama or Mistral derivatives) when on-premise inference is a hard requirement. We will architect to your data-residency and tenancy constraints and benchmark precision before committing to a model.
Let's talk about your value-based care project.
Working on a value-based care contract, ACCESS Model application, EHR integration, or AI-enabled clinical workflow project? Book a 20-minute discovery call or email [email protected].