Coding OpenMed diagnoses to ICD-10-CM / PCS

SkillDev tools

Lets your agent suggest ICD-10-CM diagnosis and procedure codes for medical conditions with explanations.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Coding OpenMed diagnoses to ICD-10-CM / PCS skill

About this capability

Suggests candidate ICD-10-CM diagnosis codes (and ICD-10-PCS procedure codes) for diagnoses and procedures extracted by OpenMed, with rationale and a human-coder caveat. Use when the user wants to code a problem list, map a diagnosis span to a billable ICD-10-CM code, route a finding to the right ch

What this skill tells your AI

The instructions your AI receives, as published by maziyarpanahi/openmed in skills/coding-icd10/SKILL.md and read by ahel’s review.

Suggest ICD-10-CM diagnosis codes (and ICD-10-PCS for inpatient procedures) for the diagnosis and procedure spans OpenMed extracts. This is decision support for a certified coder, not autonomous billing: OpenMed + this skill narrow the candidate set and explain why; a human validates the final, billable code.

ICD-10-CM and ICD-10-PCS are public domain. CMS publishes the complete annual code files, addenda, and indexes for free. (CPT/HCPCS procedure codes are AMA-licensed and restricted — out of scope here; obtain those separately under the user's own AMA license.)

When to use

  • A note yields diagnoses ("type 2 diabetes with diabetic CKD", "community- acquired pneumonia") and you want candidate ICD-10-CM codes plus rationale.
  • You need to route a span to the right chapter quickly (see references/icd10-chapters.md for code ranges).
  • You hold legacy ICD-9 codes and need an approximate GEM cross-walk.
  • You are pre-filling encounter diagnoses for a coder's review queue.

For clinical-meaning codes use mapping-to-snomed; for HCC/risk capture use coding-hcc-risk-adjustment; this skill is for the ICD-10 classification.

Quick start (public data + public FHIR lookup)

Two complementary paths, both license-clean:

A) CMS files, loaded locally (public domain; you download once):

# CMS publishes the order/addenda file; load the code->description table.
# Columns: code (no dot), description; you insert the dot for display.
icd10cm = {}                       # "E1122" -> "Type 2 diabetes mellitus with diabetic chronic kidney disease"
with open("icd10cm_order_2025.txt", encoding="latin-1") as fh:
    for line in fh:
        code = line[6:13].strip()
        billable = line[14] == "1"     # '1' = valid billable code
        long_desc = line[77:].strip()
        if billable:
            icd10cm[code] = long_desc

def search_local(term: str, limit: int = 5):
    t = term.lower()
    hits = [(c, d) for c, d in icd10cm.items() if t in d.lower()]
    return sorted(hits, key=lambda cd: len(cd[1]))[:limit]

B) A FHIR terminology server that hosts ICD-10-CM (public servers exist; e.g. an NLM Clinical Tables endpoint or your own HAPI/Ontoserver):

import requests

# NLM Clinical Tables (public, no key) — ICD-10-CM autocomplete/search:
def search_icd10cm(term: str, count: int = 7):
    r = requests.get(
        "https://clinicaltables.nlm.nih.gov/api/icd10cm/v3/search",
        params={"sf": "code,name", "terms": term, "maxList": count}, timeout=10,
    )
    r.raise_for_status()
    _total, codes, _extra, display = r.json()
    return list(zip(codes, [d[1] for d in display]))   # [(code, name), ...]

print(search_icd10cm("type 2 diabetes nephropathy"))

Workflow

  1. Extract diagnosis/procedure spans with OpenMed (Disease/Pathology models).
  2. Route to a chapter using the span's clinical theme and references/icd10-chapters.md (e.g. endocrine → E00–E89, circulatory → I00–I99). This shrinks the search space and catches obvious mis-hits.
  3. Search the code text (local CMS table or the NLM API) for candidates.
  4. Apply ICD-10-CM specificity rules in your rationale: laterality, acute/chronic, episode of care, "with"/"due to" combination codes, and "code first / use additional code" notes. Flag where the note lacks the detail a billable code requires.
  5. Rank candidates; present the top few with rationale and the missing- detail caveat, not a single auto-selected code.
  6. Emit {system: "http://hl7.org/fhir/sid/icd-10-cm", code, display} marked status: needs-coder-review, with OpenMed source offsets.

Hand-off from OpenMed

openmed.analyze_text(..., output_format="dict") returns entities, each a dict with text, label, confidence, start, end. Consume Disease/Pathology spans:

import openmed

note = "Assessment: type 2 diabetes with diabetic nephropathy; CAP."
result = openmed.analyze_text(
    note,
    model_name="disease_detection_superclinical",   # Disease category
    output_format="dict",
)

DX_LABELS = {"DISEASE", "CONDITION", "PATHOLOGY"}
for ent in result["entities"]:
    if ent["label"] in DX_LABELS:
        candidates = search_icd10cm(ent["text"], count=5)
        print(ent["text"], ent["start"], ent["end"],
              f"(conf {ent['confidence']:.2f}) ->", candidates)
        # surface as SUGGESTIONS for a coder — never auto-bill

Keep OpenMed's start/end offsets next to each suggested code so the coder can jump to the exact supporting text. Store offsets and codes only — never the raw note in your suggestion log.

Edge cases & gotchas

  • Human-in-the-loop is mandatory. ICD-10-CM coding has legal/financial weight. Output candidates with rationale; a certified coder assigns the final billable code. Never present a suggestion as an authorized claim.
  • Specificity & unspecified codes. Many billable codes demand laterality, episode, or "with" detail the note may not state. Prefer flagging "documentation insufficient for a specific code" over forcing an .9/unspecified code.
  • Combination codes. ICD-10-CM bundles related conditions (e.g. E11.22 = diabetes with diabetic CKD). Don't emit two separate codes where one combination code is required; let the search surface combinations.
  • "Code first" / "use additional code" / Excludes1/Excludes2 sequencing notes change which codes coexist. Carry these as rationale for the coder.
  • GEMs are approximate. ICD-9↔ICD-10 General Equivalence Mappings are many-to-many and lossy; treat a GEM result as a starting hint, not a billable mapping.
  • Annual updates. Codes change every fiscal year (Oct 1). Pin the file year you loaded and refresh annually; record which version produced a suggestion.
  • Licensing. ICD-10-CM/PCS are public domain (CMS). Do not pull in CPT or proprietary code maps that require an AMA/other license — those stay user-supplied and out-of-process.
  • Local-first. OpenMed NER runs on-device; if you query the NLM API, send only the de-identified diagnosis string. No PHI over the wire.

Standards & references

Signals

GitHub stars
5k
Forks
666
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
coding-icd10
Source
github.com/maziyarpanahi/openmed