maziyarpanahi/openmed/coding-hcc-risk-adjustment

Mapping conditions to CMS-HCC V28 and estimating RAF

Surface and risk-adjust the chronic conditions OpenMed extracts by mapping them to CMS-HCC categories (the V28 model, phasing in for payment years 2024–2026) and estimating a RAF (Risk Adjustment Factor) score. CMS pays Medicare Advantage plans based on RAF, so accurate, documented capture of chronic disease matters — and much of that signal lives in the narrative note, exactly what OpenMed reads.

This is decision support for coders/clinicians, not autonomous coding. The output is "candidate HCCs + estimated RAF + the documentation that supports (or fails to support) each one," for human validation.

CMS-HCC crosswalks (ICD-10-CM → HCC) and the category coefficients are public — CMS publishes them annually. Nothing restricted is bundled.

When to use

• You want to find risk-adjustable diagnoses mentioned in a note that may not

be on the coded problem list ("suspect HCCs" / recapture).

• You need to map ICD-10-CM codes to V28 HCCs and apply the hierarchy.
• You want an estimated RAF for a patient or panel for review.
• You need to check whether a diagnosis has MEAT support

(Monitored, Evaluated, Assessed, Treated) in the documentation.

Pairs with coding-icd10 (you need ICD-10-CM codes first) and may consume mapping-to-snomed output upstream.

Quick start (public CMS crosswalk + coefficients)

CMS publishes the V28 ICD-10-CM→HCC mapping and the model coefficients. Load them locally (public files) and apply the model:

import csv
# 1) ICD-10-CM -> HCC (V28) crosswalk from the CMS Risk Adjustment files.
icd_to_hcc = {}                       # "E1122" -> "HCC38" (Diabetes w/ complication)
with open("cms_hcc_v28_icd_map.csv") as fh:
    for row in csv.DictReader(fh):
        icd_to_hcc[row["icd10cm"].replace(".", "")] = row["hcc_v28"]
# 2) HCC -> RAF coefficient for the relevant model segment (e.g. CNA community).
hcc_weight = {}                       # "HCC38" -> 0.166 (illustrative)
with open("cms_hcc_v28_coefficients.csv") as fh:
    for row in csv.DictReader(fh):
        hcc_weight[row["hcc"]] = float(row["coefficient"])
# 3) Apply the HCC hierarchy: a more severe HCC in a family suppresses milder
#    ones (e.g. acute MI suppresses angina). Load the hierarchy from CMS.
hierarchy = {                         # parent HCC -> HCCs it zeroes out
    # "HCC37": {"HCC38"},  # illustrative; use the official V28 hierarchy file
}
def apply_hierarchy(hccs: set[str]) -> set[str]:
    kept = set(hccs)
    for parent in hccs:
        kept -= hierarchy.get(parent, set())
    return kept
def estimate_raf(icd_codes: list[str], demo_factor: float = 0.0) -> dict:
    hccs = {icd_to_hcc[c] for c in icd_codes if c in icd_to_hcc}
    hccs = apply_hierarchy(hccs)
    disease_raf = sum(hcc_weight.get(h, 0.0) for h in hccs)
    return {"hccs": sorted(hccs),
            "disease_raf": round(disease_raf, 3),
            "estimated_raf": round(disease_raf + demo_factor, 3)}

The demo_factor (age/sex, dual/disability, institutional status) comes from the CMS demographic tables — add it for a full RAF; omit for the disease component.

Workflow

1. Extract condition spans with OpenMed (Disease/Pathology/Oncology models).
2. Code each to ICD-10-CM (see coding-icd10) — HCCs key off ICD-10-CM.
3. Map ICD-10-CM → V28 HCC via the CMS crosswalk.
4. Apply the hierarchy so only the most severe HCC in each family counts.
5. Sum coefficients for the correct model segment + add the demographic factor

to estimate RAF.

1. Attach MEAT evidence: for each candidate HCC, cite the note text that

Monitors/Evaluates/Assesses/Treats the condition. No MEAT → flag as "unsupported / needs clinician confirmation," not a captured HCC.

1. Emit candidate HCCs + estimated RAF + supporting offsets for human review.

Hand-off from OpenMed

openmed.analyze_text(..., output_format="dict") returns entities, each a dict with text, label, confidence, start, end. Use the offsets to pull MEAT evidence sentences:

import openmed
note = ("Problem list: type 2 diabetes with diabetic nephropathy; COPD. "
        "Plan: continue metformin, ordered HbA1c, refer nephrology.")
result = openmed.analyze_text(
    note,
    model_name="disease_detection_superclinical",   # Disease category
    output_format="dict",
)
DX_LABELS = {"DISEASE", "CONDITION", "PATHOLOGY"}
suspects = []
for ent in result["entities"]:
    if ent["label"] in DX_LABELS:
        # 1) code to ICD-10-CM (coding-icd10) -> e.g. "E1122"
        icd = map_to_icd10cm(ent["text"])           # your coding step
        hcc = icd_to_hcc.get(icd)
        if hcc:
            # MEAT: capture the sentence around the span for the reviewer
            sent = note[max(0, ent["start"] - 60): ent["end"] + 80]
            suspects.append({"condition": ent["text"], "icd10cm": icd,
                             "hcc": hcc, "span": (ent["start"], ent["end"]),
                             "meat_context": sent})
raf = estimate_raf([s["icd10cm"] for s in suspects])
print(raf, suspects)        # candidates + estimate, for coder validation

Carry OpenMed's start/end offsets so every suspect HCC links to the exact documentation; this is what makes the suggestion auditable for RADV. Store codes, HCCs, and offsets — not the raw note.

Edge cases & gotchas

• Decision support, never autonomous coding. Risk-adjustment coding is

audited (CMS RADV) and has direct payment and compliance consequences. Output suspects with evidence for a certified coder/clinician; never submit HCCs automatically.

• MEAT is required. A diagnosis merely mentioned (e.g. in history) without

being Monitored/Evaluated/Assessed/Treated in the encounter generally cannot be captured. Always attach MEAT evidence and flag bare mentions as unsupported.

• V28 dropped ~2,000 codes. The V28 transition removed many ICD-10-CM codes

from HCC mapping (notably diabetes-without-complication, some vascular and inflammatory codes). A code that mapped under V24 may map to nothing under V28 — use the V28 crosswalk, not V24, and don't assume continuity.

• Hierarchy suppression. Within a disease family only the most severe HCC

counts; summing all of them inflates RAF. Apply the official V28 hierarchy.

• Model segment matters. Coefficients differ by segment (community vs

institutional, aged vs disabled, new enrollee). Use the right segment's table or the RAF is wrong.

• Negation/uncertainty. "No evidence of CHF" or "rule out malignancy" must not

become captured HCCs. Resolve assertion/negation in OpenMed before mapping.

• Annual model updates. CMS revises the model and weights yearly and is

blending V24/V28 across payment years 2024–2026; pin and record which model version and payment year your estimate used.

• Licensing. CMS-HCC crosswalks/coefficients and ICD-10-CM are public. Do not

bundle restricted vocabularies (CPT, SNOMED, UMLS) to support this — keep those user-supplied and out-of-process.

• Local-first. OpenMed NER runs on-device; HCC mapping uses local CMS tables.

No PHI needs to leave the process at all.

Standards & references

• CMS Risk Adjustment (HCC models, files, coefficients):

https://www.cms.gov/medicare/payment/medicare-advantage-rates-statistics/risk-adjustment

• CMS-HCC model software & ICD-10 mappings (V28):

https://www.cms.gov/medicare/health-plans/medicareadvtgspecratestats/risk-adjustors

• ICD-10-CM (public domain), for the upstream codes:

https://www.cms.gov/medicare/coding-billing/icd-10-codes

• RADV (Risk Adjustment Data Validation) overview:

https://www.cms.gov/research-statistics-data-and-systems/monitoring-programs/recovery-audit-program-parts-c-and-d

• Companion skills: coding-icd10 (codes feed HCCs), mapping-to-snomed.