OBJECTIVES: Length of stay (LOS) models support evaluating ICU care
however, current benchmarking models fail to consider differences in LOS between surviving and nonsurviving patients, which can lead to biased predictions toward the surviving population. We aim to develop a model addressing this as well as documentation bias to improve ICU benchmarking. DESIGN: The Critical Care Outcomes Prediction Model (CCOPM) LOS uses patient characteristics, vitals, and laboratories during the first 24 hours of ICU admission to predict LOS in the hospital and ICU using a deep learning framework for modeling time to events with competing risk. Data was randomly divided into training, validation, and test (hold out) sets in a 2:1:1 ratio. SETTING: Electronic ICU Research Institute database from participating tele-critical care programs. PATIENTS: Six hundred sixty-nine thousand eight hundred seventy-six ICU admissions pertaining to 628,815 patients from 329 ICUs in 194 U.S. hospitals, from 2017 to 2019. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: Model performance was assessed using the coefficient of determination (R2), concordance index, mean absolute error, and calibration. For individual stays in the test set, the ICU LOS model presented R2 = 0.29 and 0.23 for surviving and nonsurviving populations, respectively, at the individual level and R2 = 0.48 and 0.23 at the ICU level. Conversely, hospital LOS model presented R2 = 0.46 and 0.52 at the individual level and R2 = 0.71 and 0.64 at the ICU level. In the subset of the test set containing predictions from Acute Physiology and Chronic Health Evaluation (APACHE) IVb, R2 of ICU LOS for surviving and nonsurviving populations was, respectively, 0.30 and 0.23 for the CCOPM and 0.16 and zero for APACHE IVb. For hospital LOS, the values were R2 = 0.39 and 0.40 for the CCOPM and 0.27 and zero for APACHE IVb. CONCLUSIONS: This novel LOS model represents a step forward in achieving more equitable benchmarking across diverse ICU settings with varying risk profiles.