Answers are grounded in this paper and cite their sources.
Early warning systems for acute kidney injury are typically evaluated by discrimination, yet ranking ability does not indicate whether acting on a prediction is worthwhile, and how decision-analytic value changes with the prediction horizon has seldom been examined. This study assembled adult intensive care admissions from MIMIC-IV into a single at-risk cohort defined at a 12 h landmark, and applied four prediction windows of 6, 12, 24 and 48 h differing only in the censoring point. Discrimination, calibration and standardized net benefit were estimated for each, alongside the proportion of admissions in which the outcome could be ascertained. Ascertainment rose from 42.7% at 6 h to complete at 48 h, the latter following from the entry criteria, and varied from 14.2% to 93.4% with the hour of admission; among ascertained admissions the event proportion was similar across strata, indicating that sampling largely determined whether an outcome was observed. In the cohort ascertainable at 6 h, standardized net benefit fell by 14.9% against 5.2% for discrimination (difference 9.65% points); because the measures lack a common no-information value, this fell close to zero on a chance-corrected scale and across sensitivity analyses. Decision-analytic value is thus bounded by the measurement process rather than declining faster than discrimination.
No takes yet. Share an insight, caveat, or question.
Teng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: