Feature (Target Unit)	MIMIC-IV	AmsterdamUMCdb	eICU-CRD	HiRID
rate_epinephrine (mcg/kg/min)	mimic_derived.sofa	drugitems 6818, gamma → mcg/kg/min	infusiondrug (epinephrine) → mcg/kg/min	Givendose (pharmaid 71, 1000649, 1000650, 1000655, 1000750) → mcg/kg/min
rate_norepinephrine (mcg/kg/min)	mimic_derived.sofa	drugitems 7229 → mcg/kg/min	infusiondrug (norepinephrine) → mcg/kg/min	Givendose (pharmaid 1000462, 1000656, 1000657, 1000658) → mcg/kg/min
rate_dopamine (mcg/kg/min)	mimic_derived.sofa	drugitems 7179 → mcg/kg/min	infusiondrug (dopamine) → mcg/kg/min	constant 0.0 (absent)
rate_dobutamine (mcg/kg/min)	mimic_derived.sofa	drugitems 7178 → mcg/kg/min	infusiondrug (dobutamine) → mcg/kg/min	Givendose (pharmaid 426) → mcg/kg/min
meanbp_min (mmHg)	mimic_derived.sofa	numericitems 6642/6679/8843, hourly MAX of MAP	invasive + non-invasive MAP, hourly MIN	Observation variableid 110, 610 hourly MIN
pao2fio2ratio_novent (mmHg)	mimic_derived.sofa	PaO2/FiO2 when not ventilated	resp. P/F when mechvent = 0	Average PaO2 × 100/average FiO2 when is_ventilated = 0
pao2fio2ratio_vent (mmHg)	mimic_derived.sofa	PaO2/FiO2 when ventilated	resp. P/F when mechvent = 1	Average PaO2 × 100/average FiO2 when is_ventilated = 1
gcs_min (3–15)	mimic_derived.sofa	reconstructed from GCS components, hourly MIN	physicalexam + nursecharting GCS, MIN	Verbal GCS, Motor GCS and Eye GCS. MIN
bilirubin_max (mg/dL)	mimic_derived.sofa	numericitems 6813/9945, µmol/L ÷ 17.1	lab total bilirubin, MAX	Bilirubin MAX, µmol/L ÷ 17.1
creatinine_max (mg/dL)	mimic_derived.sofa	numericitems 6836/9941/14,216, µmol/L ÷ 88.4	lab creatinine, MAX	Creatinine MAX, µmol/L ÷ 88.4
platelet_min (×109/L)	mimic_derived.sofa	numericitems 9964/6797/10,409/14,252, MIN	lab platelets, MIN	Platelets MIN
admission_age (years)	icustay_detail.admission_age	agegroup band midpoint	patient.age (‘>89’ → 89)	General table, Age
charlson_comorbidity_index	mimic_derived.charlson	not available (left missing/uninformative)	derived from pasthistory (missing → 0)	constant 5.78 (MIMIC mean)
admission_type_Medical	rule on admission type + surgical	specialty + urgency rule	default when not surgical	constant 1.0
admission_type_ScheduledSurgical	elective & surgical	surgical & urgency = 0	operative & elective	constant 0.0
admission_type_UnscheduledSurgical	non-elective & surgical	surgical & urgency = 1	operative & emergent	constant 0.0
Discharge outcome (disch_48 h ground truth)
Discharge field (end of ICU unit stay)	icustay_detail.outtime	admissions.dischargedat	patient.unitdischargeoffset	dischargetime (via window_id)
ICU death flag (icu_expire_flag) source	Derived: dod (date-only) same day as/before outtime (vi)	Native: destination = ‘Overleden’	Native: unitdischargestatus = ‘Expired’	Native: discharge_status == ‘dead’ (Exitus) (viii)
Hospital-level mortality flag (hospital_expire_flag)	Native, independent field	Copied from icu_expire_flag (vii)	Native, independent field	Not present in schema
Death excluded from discharge/disch_48 h timeline?	No	No	No	No
Ward/step-down transfer distinguished from final discharge?	No, any destination ends the stay record identically	No, folded into “discharge alive”	No, same as MIMIC	Not fully determinable from available code
Cross-stay readmission stitching?	None, each stay modeled independently	None, each admission modeled independently	None, patienthealthsystemstayid never used	None; structurally not possible (no patient-level key released)
Notes. (i) All four vasopressor rates target mcg/kg/min (weight-normalized): eICU normalizes explicitly, Amsterdam converts via the dose-unit “gamma” rule divided by patient weight, MIMIC uses its native mimic_derived.sofa units, and the HiRID values are passed through from an upstream extract. (ii) meanbp_min is mmHg everywhere, but Amsterdam stores the hourly maximum of mean arterial pressure while eICU stores the hourly minimum—a minor cross-database aggregation difference that local retraining absorbs. (iii) A single hourly PaO2/FiO2 ratio is routed to the ventilated or non-ventilated column by ventilation status; the other column is set to 0. (iv) No database applies a sedation/intubation correction to GCS. (v) Constant-imputed cells (used only by the direct-transfer condition; local retraining re-learns from the variables that carry information): Amsterdam Charlson (missing/uninformative); HiRID dopamine rate (0), Charlson (5.78, the MIMIC training mean), and the three admission-type columns (Medical = 1, the two surgical columns = 0). (vi) MIMIC-IV’s icu_expire_flag is a date-only proxy (death on or before the calendar day of ICU outtime), coarser than the exact discharge-status fields used by the other three databases, because MIMIC-IV’s de-identified date of death carries no time-of-day. (vii) AmsterdamUMCdb’s hospital_expire_flag is copied directly from icu_expire_flag rather than computed independently, so the two are identical by construction for this database; unlike MIMIC-IV and eICU-CRD, Amsterdam cannot distinguish an in-ICU death from a later in-hospital, post-ICU death, and any hospital-level mortality figure reported for Amsterdam elsewhere in this paper should be read as identical to its ICU-level figure. (viii) In HiRID, 164 of 9150 stays (1.8%) have a missing (NaN) discharge_status and are mapped to Exitus = 0, that is, treated as survivors, a data-quality issue not present in the other three databases. (ix) In all four databases, the discharge outcome is defined and modeled per ICU stay, not per patient; none of the four ETL pipelines links the repeated ICU stays of the same individual.