Synthetic data validation · CLIF 2.1

Does the synthetic data behave like real critical illness?

A full audit of the fully-synthetic CLIF datasets against real CLIF — the ICU cohort (85,248 stays) and the whole-hospital population (365,000 stays) — across missingness, length-of-stay, life support, death, patient flow, lab-value shape, the longitudinal course of illness, and privacy.

ICU: 85,248 stays vs real ICU cohort Full hospital: 365,000 stays vs real full population Synthetic — generated, no real records

Where it should match

matched

LOS distribution (median and tails), all life support, missingness, lab-value shapes, and the illness trajectory land in the real range.

Where it differs by design

Chicago race mix, younger age, and network-median mortality — deliberate, not drift.

Lab-value distributions

shape-matched

Empirical inverse-CDF marginals reproduce skew and long tails, not just means — creatinine's kidney tail, lactate's skew.

ICU cohort — in the statistical region of the original

The ICU dataset's length-of-stay, organ support, and measurement density were fit to reproduce real CLIF ICU statistics. They do.

Metric	Synthetic	Real	Verdict
Hospital LOS — median	167 h	164 h	matched
Hospital LOS — p10 / p90 (tails)	45 / 497 h	51 / 513 h	matched
Invasive ventilation (IMV)	0.402	0.412	matched
Non-invasive ventilation (NIPPV)	0.058	0.064	matched
High-flow nasal cannula	0.061	0.069	matched
CRRT (renal replacement)	0.041	0.040	exact
Core chemistry present (creatinine, Na, Hgb)	0.985	0.986	exact
Lactate present (per stay)	0.674	0.692	matched
Arterial blood gas present (per stay)	0.690	0.695	matched

Lab values — the shape, not just the rate

Each lab is drawn through an empirical inverse-CDF fit to the real distribution, so skewed and long-tailed shapes are reproduced — a single log-normal would flatten creatinine's kidney-disease tail or lactate's skew. Values shown are the 10th / 50th / 90th percentiles.

Lab	Synthetic p10 / p50 / p90	Real p10 / p50 / p90	Verdict
Creatinine (mg/dL)	0.53 / 1.0 / 2.86	0.5 / 1.0 / 3.2	tail matched
Lactate (mmol/L)	0.9 / 1.7 / 4.41	0.9 / 1.8 / 4.8	matched
Sodium (mmol/L)	132 / 138 / 144	132 / 138 / 145	exact
BUN (mg/dL)	8 / 19 / 55	9 / 22 / 63	matched
Platelets (10⁹/L)	74 / 205 / 384	64 / 193 / 393	matched

The longitudinal story — deterioration toward death

The hardest test: does a dying synthetic patient get sicker the way a real one does? Aligned to the last 48 hours before death or discharge, decedents should diverge from survivors — and they do, in both datasets.

Decedents Survivors Synthetic Real CLIF

Different — by design, not by accident

The dataset is deliberately a Chicago-representative, network-median cohort — statistically distinct from its source cohort so no one dataset can be mistaken for the other. CI also locks those network-median rates (LOS, IMV, vaso, CRRT, ABG co-occurrence) against the shareable base_pack on every PR — no DUA required to run the gate.

Characteristic	Synthetic	Real CLIF	Rationale
Race — White / Black	42% / 30%	67% / 9%	Chicago, not Boston
Age — median	63	67	shifted younger
In-hospital mortality	0.095	0.115	CLIF network median
Female	53%	43%	cohort shift

The whole hospital — realistic patient flow

The second dataset is a full hospital population (365,000 stays): most patients never leave the ward, ~15% reach the ICU, and they arrive through a realistic mix of front doors. The population statistics match real; the flow is a deliberately cleaner design (more planned ICU access, fewer floor deteriorations) than the raw source cohort.

Metric	Synthetic	Real full pop.	Verdict
Reach the ICU	0.150	0.156	matched
Hospital LOS — median	64 h	67 h	matched
Hospital LOS — p10 / p90	13 / 226 h	13 / 243 h	matched
In-hospital mortality	0.022	0.021	matched
Direct admissions	0.20	0.06	by design
Arrive via ER / OR	0.55 / 0.18	0.84 / 0.10	by design
Transferred into ICU (deteriorate)	0.093	0.139	by design <10%
Direct / planned ICU admit	0.057	0.017	by design

Front-door mix (synthetic): ED 57% · direct-to-ward 21% · OR/procedural 11% · direct-ICU 6% · stepdown 6%.

Privacy — no record traces back

The generator samples from aggregate parameters and never copies a real record. Three standard privacy measures (synthetic vs real, 4,000 each) confirm no synthetic patient is a stand-in for a real one.

Measure	Value	Reading
Distance to closest real record — median (p5)	3.49 (1.89)	no copies even the closest 5% sit ~1.9σ from any real record
Nearest-neighbour distance ratio — median	0.97	no unique linkage ≈1 means not singled-out to one real record
Identifiability	0.25	moderate the fidelity/privacy trade-off of matching real value shapes

Honest gaps

Arterial-gas panel union runs slightly high — 0.76 vs 0.70. Real blood gases are co-ordered as a panel, and presence is drawn through a fitted Gaussian copula so co-ordered labs appear together (this corrected an earlier several-fold inflation). One residual remains: SₐO₂, measured less consistently (0.36), still contributes semi-independently, so "any arterial gas present" sits a few points above real.

White-cell counts skew a touch low. WBC p90 is 14 vs 18 real — the empirical marginal reproduces most of the distribution but slightly under-weights the high-count (leukocytosis) tail. Every other lab's value distribution tracks real, including the tails.

Verdict

Both datasets reproduce real CLIF where it counts — missingness, length-of-stay distributions, life support, death, lab-value shapes, and the death-and-recovery trajectory — the ICU cohort against the real ICU population and the full-hospital dataset against the real whole-hospital population. They are intentionally distinct where designed to be: Chicago demographics, network-median mortality, and a cleaner patient flow. And they carry no traceable records. Statistically similar, believably real, but different and safe to share.