MAIP — Not all ICU patients with the same diseases cost the same

Not all ICU patients with the same chronic diseases cost the same

We identified 9 distinct patient types within comorbidity groups. Their hospital costs differ by up to 2.5× and you can tell which type at admission.

Hari S. Sreedeth · MSc Health Data Science, UNSW · Portfolio project

In a typical ICU, patients are grouped by their chronic conditions., such as heart failure, COPD, diabetes. But within those groups, costs and outcomes vary enormously. Two patients with the same comorbidity burden can have median hospital costs of $9,000 and $23,000.

Using the SUPPORT-II cohort (9,105 ICU admissions), I built an unsupervised learning pipeline that first stratifies patients by chronic disease burden, then discovers distinct acute phenotypes within each group, separating, for example, a low-acuity cancer patient from a neurologic catastrophe patient, even when both have zero recorded comorbidities.

9,105

ICU admissions analysed
across 3 comorbidity strata

Distinct phenotypes
identified (3 per stratum)

2.5×

Cost difference within
the same comorbidity level

The finding: same comorbidity, different costs

Phenotype

Median LOS

Median cost

Relative cost

Low comorbidity (0–1 chronic conditions, n = 4,183)

Low-acuity solid tumourCancer patients with preserved organ function

5 days

$11,353

Neurologic catastrophe / comaMarked neurological compromise, moderate systemic derangement
8 days
$21,114

Acute multi-organ failureSevere acute illness despite low baseline chronic disease
4 days
$9,207

Mid comorbidity (2–3 chronic conditions, n = 3,822)

Older COPD-dominant respiratoryCOPD with moderate heart failure and diabetes

6 days

$12,189

Multi-organ failure + malignancyAcute renal failure, metastatic cancer, cirrhosis
9 days
$22,042

Compensated heart failure + tumoursStable cardiac disease with solid tumour burden

7 days

$14,842

High comorbidity (4+ chronic conditions, n = 1,100)

Diabetic multi-organ failureRenal failure, diabetes, cancer, highest acuity
9 days
$23,412

Cardio-pulmonary + liver diseaseHeart failure, COPD, cirrhosis, intermediate severity

7 days

$17,358

Cardiometabolic + solid tumoursHeavy chronic burden but least acutely deranged

7 days

$18,362

Median hospital length of stay and total medical costs for 9 ICU phenotypes discovered by multi-view unsupervised learning (SNF-lite) within multimorbidity strata. No outcome data was used to define the phenotypes — costs and mortality were held out for validation only. Differences within each stratum are statistically significant (Kruskal–Wallis p < 0.05 for high stratum; p < 0.001 for mid and low).

Why this matters for healthcare analytics

Standard risk adjustment groups patients by comorbidity count or DRG. This analysis shows that within the same comorbidity tier, acute presentation patterns create 2–3× cost variation that comorbidity-based models miss. The phenotypes are identifiable from admission-day data (vitals, labs, demographics) and can be assigned using a simple decision tree — no complex model needed at the point of care.

Each phenotype was independently validated against 365-day mortality (Cox PH, adjusted for age, sex, comorbidity count, and acute severity). Hazard ratios ranged from 0.51 (protective) to 2.36 (high risk) relative to stratum references. A surrogate decision tree distils the phenotype assignments into auditable rules with 3–4 conditions each.

Built on the SUPPORT-II cohort (Vanderbilt University). Pipeline includes multi-view similarity network fusion, bootstrap stability validation, cross-validated prognostic assessment, and LLM-assisted rule translation with programmatic QC.

Full methodology → GitHub →