Extract Recform Icd10 Panels

Extract — Requisition Form → panel/ICD-10 JSON

Give it a requisition PDF; it returns one JSON mapping each test panel to its ICD-10 code lists. Runs standalone, and feeds import-lab-recform Workflow 05 (icd10-panel-fill / AnnotsJSON) so the panel→code mapping is derived from the form instead of typed by hand.

Output shape (this is the contract)

{
  "relevant_diagnosis_condition": {
    "HEREDITARY PERIPHERAL NEUROPATHY": {
      "PRIMARY ICD-10 CODES":   { "icd_codes": ["G60.0"], "required": true },
      "SECONDARY ICD-10 CODES": { "icd_codes": ["Z82.0", "M62.81", "R20.2"], "required": false }
    },
    "HEREDITARY PERIPHERAL NEUROPATHY NGS PANEL": {
      "PRIMARY ICD-10 CODES":    { "icd_codes": ["…"], "required": true },
      "SECONDARY ICD-10 CODES":  { "icd_codes": ["…"], "required": false },
      "CROSS-PANEL ICD10 CODES": { "icd_codes": ["…"], "required": false }
    }
  }
}

• icd_codes holds code strings only — no descriptions.
• required: true for PRIMARY, false for SECONDARY and CROSS-PANEL.
• Also keep source_pdf, form_title, lab, extracted_at, extraction_method, and a notes array for every judgement call the reader had to make. The validator ignores those keys.

Panel naming

Each disease group on the form carries TWO ICD tables:

A group printed as NO PCR / MLPA / REPEAT EXPANSION (e.g. HEREDITARY DEMENTIA) gets only the … NGS PANEL entry. Copy the group heading verbatim, including & and any parenthetical.

Flow

1. Check the text layer FIRST

pdftotext -layout "<pdf>" - | head -40

Real text → parse it, do not read images. Empty → flattened form, continue. (The AlphaDERA neuro form: 4 pages, 1 image each, 0 characters.)

2. Render + slice

.claude/skills/extract-recform-icd10-panels/scripts/render-recform-pages.sh "<pdf>" <out-dir> [first] [last]

300 DPI, three ~1700px-wide strips per page with 80px overlap, and it prints the text-layer character count so step 1 is recorded.

Never read codes off the page-level render. At page scale the digits are genuinely ambiguous: a first pass on 2026-07-31 misread M62.81 as M82.81, G60.2 as G80.2, and F02.B0 as F02.80. Read the strips.

3. Read every strip with the Read tool

Per group collect: the PRIMARY column, the SECONDARY column — it has two sub-columns, the right one is easy to miss — and CROSS-PANEL ICD10 CODES when present. Tables span pages: the HEREDITARY EPILEPSY NGS table starts at the bottom of page 2 and its codes are at the top of page 3.

4. Write the JSON, then validate

python3 .claude/skills/extract-recform-icd10-panels/scripts/validate-recform-json.py <json>

Checks: every panel has a PRIMARY block, icd_codes non-empty, required matches the block type, no duplicates inside a list, and every code matches ICD-10-CM shape (letter-suffix codes G40.A, G40.3A1, F02.A0, G20.A1 and the form's .x family placeholders included). It prints a per-panel count table — compare those counts against the strips before declaring done. Exit 1 on errors.

The validator catches malformed codes, never a misread one. Anything ambiguous goes in notes, not silently into icd_codes.

5. After the codes are imported — audit them

verify-recform-icd10-import diffs the dashboard AnnotsJSON against this JSON (unfilled entries, codes never imported, wrong panel, extras, count mismatches), and its Phase B does an independent second read of the PDF to catch misreads that no format check can see. Run it after import-lab-recform W05.

Transcription rules (decided 2026-07-31 — keep consistent)

• I42.1/I42.2 (Hypertrophic cardiomyopathy) is one printed row for two codes → split into I42.1 and I42.2.
• A CROSS-PANEL row printed as a combination (G60.0 (…) + G71.09 (…)) → split into separate entries in the same list.
• F80.x, Z80.x, C50.x, C64.x stay exactly as printed — family placeholders, not billable codes. Never invent a 4th character.
• Gene lists (GJB1, MPZ, MFN2 …), methodology checkboxes and EXTENDED PANEL gene lists are NOT part of this output — ICD tables only.

Done: Neurological Disorders — AlphaDERA Labs

data/neurological-disorders-alphadera-labs.json — 15 panels, 35 blocks, 128 distinct codes, validator clean. Groups: Hereditary Peripheral Neuropathy, Hereditary Ataxia & Hereditary Spastic Paraplegia, ALS & Motor Neuron Disease, Hereditary Epilepsy & Epileptic Encephalopathy, Neurodevelopmental Panels (XLID / Autism / Migration), Hereditary Movement Disorders, Hereditary Dementia, Hereditary Neuromuscular Disorders.

Report

✅ extract-recform-icd10-panels done
   PDF:        <file>  (<n> pages, text layer: <n> chars)
   Panels:     <n>  (PRIMARY <n> / SECONDARY <n> / CROSS-PANEL <n> blocks)
   Codes:      <n> distinct
   Output:     .claude/skills/extract-recform-icd10-panels/data/<name>.json
   Validator:  OK (0 errors)
   Ambiguous:  <list, or none>