MIMIC-IV `predicates.yaml` lab codes are `LAB//<itemid>//<unit>`, but the pinned ETL now emits `LAB//RESULT//<itemid>//<unit>`

Summary

src/MEDS_DEV/datasets/MIMIC-IV/predicates.yaml references labevents lab codes as
LAB//<itemid>//<unit>, but the pinned ETL emits measured lab results as
LAB//RESULT//<itemid>//<unit> (with the value-less order event as
LAB//SPECIMEN_COLLECTED//<itemid>). On a freshly built MIMIC-IV MEDS dataset the labevents
predicates therefore match zero events.

This is an internal inconsistency on main: datasets/MIMIC-IV/requirements.txt pins
MIMIC-IV-MEDS==0.1.0 (the release that produces LAB//RESULT//), while predicates.yaml
still uses the older LAB//<itemid> scheme. Only the labevents itemids (5xxxx) are affected;
chartevents itemids (22xxxx) still emit LAB//<itemid>//<unit> and are fine.

Evidence

creatinine_1 → LAB//50912//mg/dL. In the built data:

LAB//50912//mg/dL               -> absent (0 events)
LAB//RESULT//50912//mg/dL       -> present  (the measured result)
LAB//SPECIMEN_COLLECTED//50912  -> present  (value-less order event)

Reproduce:

import glob, polars as pl
files = glob.glob('<mimic_meds_root>/data/**/*.parquet', recursive=True)
codes = set()
for f in files[:30]:
    codes |= set(pl.scan_parquet(f).select('code').unique().collect().to_series())
print('LAB//50912//mg/dL' in codes)          # False
print('LAB//RESULT//50912//mg/dL' in codes)  # True

Impact on the registered tasks (all supported_datasets: [MIMIC-IV])

Because creatinine = or(creatinine_1, creatinine_2) and both members are labevents:

• abnormal_lab/blood_chemistry/elevated_creatinine — creatinine and
  abnormally_high_creatinine both resolve to labevents-only → match nothing → the target
  window has: creatinine (1, None) is never satisfied → empty / degenerate cohort.
• hyponatremia (sodium), metabolic_acidosis (bicarbonate), anemia (hemoglobin),
  leukocytosis (wbc), thrombocytopenia (platelets) — each or(...) keeps its surviving
  chartevents member, so the task still runs but silently drops the labevents
  contribution (biased cohort / shifted label prevalence).
• hypotension (MAP), mortality/in_icu — MAP is all chartevents; mortality is structural →
  unaffected.

Proposed fix

Insert the RESULT// infix on the 8 labevents codes (each appears twice — a plain and an
abnormally_* variant); leave the chartevents (22xxxx) codes as-is. E.g.:

  creatinine_1:
    code: LAB//RESULT//50912//mg/dL
  creatinine_2:
    code: LAB//RESULT//52546//mg/dL
  abnormally_high_creatinine_1:
    code: LAB//RESULT//50912//mg/dL
    value_min: 1.3
    value_min_inclusive: False
    value_max: null
  # ...and likewise abnormally_high_creatinine_2, sodium_2/3, abnormally_low_sodium_2/3,
  # bicarbonate_2, abnormally_low_bicarbonate_2, hemoglobin_2, abnormally_low_hemoglobin_2,
  # wbc_2, abnormally_high_wbc_2, platelets_2, abnormally_low_platelets_2.

Full list of code strings to change (16 predicate entries, 8 distinct codes):
LAB//50912//mg/dL, LAB//52546//mg/dL, LAB//50983//mEq/L, LAB//52623//mEq/L,
LAB//50882//mEq/L, LAB//50811//g/dL, LAB//51300//K/uL, LAB//51265//K/uL
→ each prefixed with RESULT//.

Longer term: a check that catches predicate ↔ ETL drift

This class of bug (a predicate silently matching zero events after an ETL code-schema change)
is invisible today — the task still "runs," just on an empty or biased cohort. Some options to
catch it in CI:

• Vocabulary check. Validate predicate codes against a reference code vocabulary for the
  pinned ETL version. (Note: metadata/codes.parquet is currently incomplete — it only lists
  codes that got a metadata-crosswalk join and omits chartevents vitals / structural / death
  codes — so this check should scan the data vocabulary, not codes.parquet.)
• Couple the pin to the predicates. When requirements.txt bumps a dataset's ETL pin, require
  the predicate-validation job above to pass in the same PR, so the code schema and the predicates
  can't drift apart again.

Happy to send a PR for the predicate fix once the
direction is confirmed.

(generated with Claude, reviewed manually)

Start with src/MEDS_DEV/datasets/MIMIC-IV/predicates.yaml and compare its labevents codes with the pinned MIMIC-IV-MEDS==0.1.0 output described in datasets/MIMIC-IV/requirements.txt. Verify the listed 16 predicate entries use RESULT// for labevents while chartevents codes remain unchanged, then reproduce the code-vocabulary scan to confirm the predicates match measured results.