ICU readmission

Unplanned ICU return during the same hospitalization End of first ICU stay

Potential use: discharge readiness assessment. We prioritize calibrated risk at ICU discharge and the tradeoff between missed returns and unnecessary continued ICU care.

Retrospective benchmark results do not establish benefit in clinical care.

Models

External AUROC is the mean AUROC across eligible external sites, weighted by test hospitalizations. MIMIC-CLIF remains a public reference. Full-report selections do not change this summary.

Task definition and ranking methods

Endpoint: unplanned return to the ICU during the same hospitalization. Prediction time: end of the first ICU stay.

Endpoint

Unplanned ICU return during the same hospitalization

Prediction time

End of first ICU stay

Reporting mode

Episodic episodic

Task threshold

0.20

High-risk alarm rule

Score ≥0.20

Decision-curve action

Continued ICU care, score ≥0.20

Potential action

discharge readiness assessment evaluated retrospectively; not a clinical recommendation

Task ID

icu_readmission

Task class

outcome

Temporal pattern

episode

Prediction unit

one-per-stay

Task type

binary

Label column

label_icu_readmission

The external summary weights each eligible site by test hospitalizations, never daily prediction rows. Scores and intervals in Summary use two decimals; ordering uses the available source precision before display rounding. Full report retains more precision. Formal ranks are pending verification of identical task versions, evaluation protocols, eligible site sets and per-site test partitions across submissions. A change requires a new ranking version and reference-model reevaluation. Exact ties must share a rank once comparability is verified. The current exports do not establish all these conditions. Only per-site confidence intervals are shown in Summary; averaging site confidence limits does not establish an interval for the ranking score.

Summary policy: flair-manuscript-2026-10-18. A missing required external estimate leaves the overall summary unavailable; remaining site reports stay accessible.

Which result each site publishes: the site a model was trained on publishes its local arm; every other site publishes external validation. A site's own-data result is never presented as external validation.

Intervals: 95% percentile-bootstrap intervals (500 replicates specified) resample hospitalizations at each site. They are not adjusted for multiplicity across tasks, sites, or thresholds, no formal hypothesis tests are performed, and they should not be used to infer statistically significant differences between sites.

Validation tags: external validation means a model is evaluated on participating external site data. Local validation means the model is evaluated on the held-out test split of the site it was trained on.

Want your model evaluated? Request a run →