docs(mimic-iv): state there is no official itemid→LOINC map

Answer #1972 without promising a maintained crosswalk in this repo.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: Chessing234

mimic-iv/docs/LOINC_MAPPING.md

@@ -0,0 +1,40 @@
+# MIMIC-IV itemid → LOINC mapping
+
+There is **no official, complete MIMIC-IV `itemid` → LOINC map** in this
+repository or on PhysioNet as a maintained product.
+
+See [issue #1972](https://github.com/MIT-LCP/mimic-code/issues/1972).
+
+## What exists today
+
+| Path | Scope | Caveat |
+|------|-------|--------|
+| `d_items` / `d_labitems` labels | MIMIC-IV dictionaries | Human-readable names, not LOINC |
+| MIMIC-III `conceptid` / community maps | Older releases | Incomplete; IV itemids are a different namespace |
+| OMOP / Athena / manual review | External | Best for modeling when you need LOINC semantics |
+| PhysioNet forum threads | Community | Useful pointers; not version-guaranteed |
+
+A multi-hop IV → III → LOINC conversion is **error-prone**: itemids were
+reassigned across CareVue / MetaVision / IV, and LOINC bindings in older
+helpers were never exhaustive.
+
+## Recommended practice for predictive modeling
+
+1. Prefer **MIMIC-IV `itemid`** (and `d_labitems.label` / `fluid` / `category`)
+   as the primary feature key inside a single MIMIC version.
+2. If LOINC is required for multi-site harmonization, build an explicit map
+   for the **subset of itemids you use**, with manual clinician/terminology
+   review — do not treat a scraped IV→III→LOINC chain as ground truth.
+3. Record the MIMIC version, map provenance, and unmapped rate in the paper
+   or model card.
+
+## What this repository will not add as “official”
+
+A full LOINC crosswalk would need ongoing clinical terminology maintenance
+outside the CSV loaders. Until PhysioNet publishes one, please treat any
+community CSV as **unofficial**.
+
+## Related
+
+- Lab value / unit quirks: `mimic-iv/buildmimic/TABLE_NOTES.md` (when present)
+- Qualitative labs (`NEG`/`POS`): same notes / issue #1938