Development and validation of a prediction model of length of ICU stay for benchmarking: a retrospective cohort study.

Efficient use of intensive care unit (ICU) resources requires accurate prediction of length of stay (LOS), yet existing LOS models show inconsistent performance across settings and may not generalize to Japanese ICUs. We conducted a retrospective cohort study using the Japanese Intensive Care Patient Database, including adults admitted to 87 ICUs between April 2022 and March 2023. The primary outcome was ICU LOS in days. Generalized additive models with several distributional assumptions were fitted using the Acute Physiology and Chronic Health Evaluation III score, primary disease category, emergency surgery status, and admission source as predictors. Model performance was evaluated using deviance explained with internal validation. The best-performing model was applied to derive ICU-level standardized length-of-stay ratios (SLOSR) and observed minus expected length of stay (OMELOS), summarized using funnel plots with and without overdispersion correction. Among 65,395 patients, median ICU LOS was 3 days (interquartile range: 2-5). A gamma model with a log link achieved the highest cross-validated deviance explained (0.415). Overdispersion correction reduced the proportion of ICUs exceeding the 95% control limits from 69% to 9.2%. These findings indicate that a gamma-based generalized additive model enables case-mix-adjusted ICU LOS benchmarking and that overdispersion correction substantially affects outlier identification.