Photo Credit: xavierarnau
The following is a summary of “Prediction of Intensive Care Length of Stay for Surviving and Nonsurviving Patients Using Deep Learning,” published in the February 2025 issue of Critical Care by Brochini et al.
Researchers conducted a retrospective study to develop a model addressing differences in length of stay (LOS) between individuals who survived and non-survived and documentation bias to enhance ICU benchmarking.
They developed the Critical Care Outcomes Prediction Model (CCOPM) LOS using data on characteristics, vitals, and laboratory results collected during the first 24 hours of ICU admission. The model predicted hospital and ICU stay durations using a deep learning framework for time-to-event modeling with competing risks. Data was split into training, validation, and test (holdout) sets in a 2:1:1 ratio. The study utilized the Electronic ICU Research Institute database (from participating tele-critical care programs), analyzing 6,69,876 ICU admissions involving 6,28,815 individuals from 329 ICUs across 194 U.S. hospitals (2017–2019).
The results showed that model performance was evaluated using the coefficient of determination (R2), concordance index, mean absolute error, and calibration. For ICU LOS, the model achieved R2= 0.29 for surviving and 0.23 for nonsurviving individuals at the individual level, and R2 = 0.48 and 0.23 at the ICU level. For hospital LOS, R2 was 0.46 and 0.52 at the individual level and 0.71 and 0.64 at the ICU level. In a test subset with Acute Physiology and Chronic Health Evaluation (APACHE) IVb, ICU LOS R2 was 0.30 for survivors and 0.23 for nonsurvivors with the CCOPM, compared to 0.16 and 0 for APACHE IVb. Hospital LOS R2 values were 0.39 and 0.40 for the CCOPM, compared to 0.27 and 0 for APACHE IVb.
Investigators concluded that the new LOS model advances equitable benchmarking across diverse ICUs with varying risk profiles.
Source: journals.lww.com/ccmjournal/abstract/9900/prediction_of_intensive_care_length_of_stay_for.456.aspx
