Add INSPIRE dataset — task-contract predicates + INSPIRE-MEDS 0.1.0

[florian6973]

Add INSPIRE dataset — task-contract predicates + INSPIRE-MEDS 0.1.0

Registers the INSPIRE (VitalDB / SNUH) dataset. This builds on the work in #308 (its registration commits are included and preserved) and adds two corrections so INSPIRE works end-to-end against the shared tasks and the current PhysioNet layout.

Contents

• INSPIRE dataset registration — README.md, dataset.yaml, refs.bib (from Add the INSPIRE dataset #308).

• predicates.yaml — rewritten to the MEDS-DEV task contract. The original predicates matched zero events and could not drive any shared task:

  • codes had unit and name reversed (the ETL emits LAB//<name>//<unit>, e.g. LAB//creatinine//mg/dL, not unit-first);
  • structural predicates used ^X//.* regexes (copied from MIMIC-IV, whose codes carry a // suffix), but INSPIRE emits bare scalar codes (HOSPITAL_ADMISSION, ICU_ADMISSION, …);
  • predicate names (creatinine_mgdl, map_mmhg, …) didn't match the names the tasks reference, and there were no abnormally_* thresholds.

  This version uses the task-contract names + adds the threshold predicates (creatinine>1.3, sodium<135, bicarbonate<22, hemoglobin<13, wbc>11, platelets<150; MAP<65), with units transferred from MIMIC-IV (INSPIRE's native units are numerically equivalent). Verified against a full INSPIRE 1.4.2 build (data vocab + value distributions): 7 of 8 shared tasks resolve through ACES with non-degenerate thresholds — all except any ED-based task (INSPIRE has no ED encounters, so ED_registration/ED_discharge are intentionally omitted).

• requirements.txt — INSPIRE-MEDS==0.1.0 (was 0.0.13). 0.1.0 adds native reading of gzipped raw tables, so the current PhysioNet 1.4.x layout (operations.csv.gz, labs.csv.gz, …) builds with no manual decompression. Drop-in: same meds-transforms~=0.2.4 pin and a byte-identical event_configs.yaml, so the predicates above are unaffected.

Notes

• Remaining step to advertise task support: add INSPIRE to each supported task's metadata.supported_datasets.
• Happy to fold these two commits into Add the INSPIRE dataset #308 instead if the maintainers prefer — whatever's easier to review.

🤖 Generated with Claude Code