The reference baseline

An optional XGBoost reference model over point-in-time features that runs the full model-to-data loop.

flair-baseline is an optional reference model: XGBoost over point-in-time features built only from events at or before each row's feature_cutoff_dttm - min / max / mean / median of each numeric code (labs, vitals, respiratory values), event counts for medications, ever-occurred flags for respiratory device and mode, plus age, sex, and hours since the hospitalization's first event. The MIMIC-CLIF reference models use an input vocabulary of 213 codes, which yields 448 inputs; XGBoost routes missing numeric values natively, and absent medication and categorical events are encoded as zero. It's the easiest way to see a report bundle without writing your own model - and it's what ships frozen to sites for external validation. For each task it runs: task.build() → Table 1 → CLIF→ELF event ETL → point-in-time featurizer → XGBoost (fit or score) → report bundle.

Prepare, then score

Prepare the data once per site. prepare chains three stages: build-cohorts (one cohort + Table 1 per task), build-data (one shared CLIF→ELF event store over every task cohort) and featurize:

flair-baseline prepare \
  --clif-config clif_config.json \
  --out . \
  [--task icu_daily_mortality]   # default: all 4 tasks

Splits come from the task cohorts: on MIMIC-CLIF a reproducible 75/25 assignment at the stay level (a stitched hospitalization never spans train and test); at other sites the oldest 75% for training and the newest 25% for testing unless explicit calendar cutoffs are given (--train-end / --test-start).

Train at the source site (MIMIC-CLIF), producing the models that ship:

flair-baseline local-training \
  --clif-config clif_config.json \
  --out .

Score at another site. Three model commands share the same local 25% holdout:

flair-baseline external-validation --models-dir mimic_baseline_models --clif-config site_clif_config.json --out .
flair-baseline transfer-learning   --models-dir mimic_baseline_models --clif-config site_clif_config.json --out .
flair-baseline local-training      --clif-config site_clif_config.json --out .

external-validation scores the shipped model, frozen, on that site's holdout; transfer-learning continues training the shipped model on the site's training split; local-training fits a new model on it. Training commands run an Optuna search by default (--hpo-trials 30; --no-hpo for fixed parameters). The report mode per task comes from its META (continuous for the mortality task and discharge to LTACs, episodic for extubation failure and ICU readmission).

Output: three site-prefixed folders

Every run writes under --out into three folders prefixed with site from the config. The PHI folder never leaves the site.

<site>_baseline_phi/_shared/MEDS/                  shared event store                              ← stays local (PHI)
<site>_baseline_phi/<task>/                       cohort.parquet, features.npz, preds_<kind>.parquet  ← stays local (PHI)
<site>_baseline_non_phi_for_upload/<task>/        codes.parquet, table1.json, report/<kind>/*.json (+viz)  ← upload this
<site>_baseline_models/<task>/<kind>/             model.json, vocab.json, params.json

<kind> is local, external_validation or transfer, one per model command. The MIMIC-CLIF models ship as mimic_baseline_models/<task>/local/; the PHI and non-PHI folders carry the running-site prefix.

flair_elf_config.yaml - the feature source of truth

This YAML (--elf-config on prepare / build-data) lists the feature domains extracted into ELF, and is the single source of truth for what the model may use (no separate exclude list):

VITAL, LAB, RESP, PA (GCS + RASS), MED_CON, MED_INT