Healthcare pack
Claims line items CSV import API
Import healthcare claim line items from CSV with CPT and ICD-10 code validation. FHIR Claim.item-shaped.
curl -X POST https://api.adaptivmapr.com/v1/uploads \
-H "Authorization: Bearer $ADAPTIVMAPR_API_KEY" \
-F "template=claims_line_items_v1" \
-F "file=@your_data.csv"Canonical columns
The whole schema, printed as it ships.
Every canonical column, the type each row carries, whether it is required, the field-level validators that fire on commit, and the multilingual header hints the cascade resolves against. This is the shipped definition, not a summary of it.
claims_line_items_v1FHIR · Claim.item- fields
- 4
- required
- 2
- validated
- 2
- hints
- 12
| Canonical column | Type | Required | Validators | Header hints the cascade matches |
|---|---|---|---|---|
cpt_code | string | yes | cpt_code | cptprocedure |
icd10_code | string | — | icd10_code | icdicd-10diagnose |
amount | number | yes | — | betragmontantimporteamount |
service_date | date | — | — | leistungsdatumdate du servicefecha del servicio |
Read the same definition as JSON at GET /v1/templates/claims_line_items_v1. A hint match resolves on layer 2 — no LLM call, no token spend, just the flat per-map fee. Hover a validator id to see what it checks.
- 4 canonical fields
- 2 required
- 2 validated
- 12 header hints, 5 languages
- FHIR Claim.item
Why it exists
Written for the file you actually receive.
The Claims line items template is the schema for billing claim detail — each row is one procedure performed for one patient on one day, priced and coded. The template is `medium` risk because it indirectly carries PHI (a CPT code tied to a service date and a patient id is identifying), so full-data mode is gated behind the AdaptivMapr subscription. CPT codes are validated against the 5-digit shape; ICD-10-CM codes against the A00 / A00.0 pattern; amounts against locale-aware number parsing. The template emits as FHIR Claim.item, which is the standard cross-border representation for billing detail. Customers use it when porting between practice-management systems, when feeding a clearinghouse, or when building an internal denial-analysis pipeline. The cascade picks up multilingual variants (Leistungsdatum, date du service, fecha del servicio) without an LLM call.
The amount field is currency-agnostic — claims systems usually run single-currency per workspace, so there is no currency column in the canonical shape. Service_date is required and parsed across the same date formats as the rest of the catalogue. CPT codes resolve via the same fallback resolver as LOINC and ATC; a "99213" stays a "99213" but a "office visit established 15min" resolves to "99213" when confidence is high enough.
Migration scenarios & the foreign headers they ship
Migration scenarios for Claims line items: feeding a clearinghouse with a freshly extracted batch, building an internal denial-analysis pipeline that runs on a year of historical line items, porting between practice-management systems where the line-item shape differs subtly, and reconciling payer remittance advices against billed lines. Foreign headers we see in the field: "CPT / Code CPT / Code de procédure / Codigo CPT / ICD / ICD-10 / Diagnose / Diagnostic / Diagnóstico / Betrag / Montant / Importe / Importo / Leistungsdatum / Date du service / Fecha del servicio / Data del servizio". Because billing files are the most regulated CSV input in healthcare, validator coverage matters more here than anywhere else — both CPT and ICD-10 run through canonical-format validators with resolver fallback for free-text descriptions.
The cascade
Five layers, and the cheapest one wins.
Layers run in order and stop the moment a column resolves. That is the single biggest cost lever in the system: a column caught on layer 2 never reaches the metered layer 5.
- L1Statisticsno LLM
Auto-accepts a header that past confirmations already resolved the same way, at {minN:100, minRatio:0.95} or {minN:20, minRatio:1.00}.
- L2Heuristicno LLM
Normalises accents, punctuation and whitespace, then compares against the column name, the label, and every registered hint (DE / FR / IT / EN / ES).
- L3Fuzzyno LLM
Token-set ratio plus Levenshtein over the normalised strings. Auto-accepts at 0.80 — it absorbs typos and reordered words.
- L4Semanticcheap, cached
Embedding cosine between the header and the field’s label + hints. Catches the long tail of paraphrases.
- L5LLMmetered
Everything still unresolved goes up in ONE batched, collision-aware call, constrained to this template’s column set so it cannot invent a field.
Try it
One template id, two ways in.
REST for your import pipeline, MCP for your editor. Both run the same cascade and both honour the same schema-only clamp.
REST · POST /v1/uploads
Name the template; the cascade picks up the rest. The canonical definition is read-only at GET /v1/templates/claims_line_items_v1.
curl -X POST https://api.adaptivmapr.com/v1/uploads \
-H "Authorization: Bearer $ADAPTIVMAPR_API_KEY" \
-F "template=claims_line_items_v1" \
-F "file=@your_data.csv"MCP · Cursor / Claude Desktop
Drop AdaptivMapr into your editor and call the same cascade as a tool. Schema-only calls leave only column names and up to three clamped sample rows.
// In Cursor or Claude Desktop with the AdaptivMapr MCP server installed:
adaptivmapr.match_headers({
template_id: "claims_line_items_v1",
headers: ["cpt_code", "icd10_code", "amount", "service_date"]
})The mappings response comes back flagged. PATCH /uploads/:id/mappings returns requires_hitl: true and hitl_status: "pending_review" so you can hold the commit in your own workflow — the flag is a signal, not a queue we run. Schema-only mode (headers plus at most three sample rows, each clamped to 80 characters) is a data-minimization mode enforced at the HTTP edge. Full-data mode routes the metered layer-5 call to phi-cloud in-region under a BAA, costs 20% more on the whole map charge, and stays locked until the workspace accepts the BAA/NDA in Settings → Security & Data.
Questions
Claims line items CSV import — FAQ
Ready when you are
Map claims line items in production — without shipping raw records.
Schema-only mode leaves only headers and a handful of clamped samples. Add full-data when you need row-level AI, routed in-region under a BAA.