A literature-based approach to predict continuous hospital length of stay in adult acute care patients using admission variables: A single university center experience.

Abstract

To review the existing literature on predicting length of stay (LOS) and to apply the findings on a Real World Data example in a single hospital. Performing a literature review on PubMed and Embase, focusing on adults, acute conditions, and hospital-wide prediction of LOS, summarizing all the variables and statistical methods used to predict LOS. Then, we use this set of variables on a single university hospital and run an XGBoost model with Survival Cox regression on the LOS, as well as a logistic regression on binary LOS (cut-off at 4 days). Model metrics are the concordance index (c-index) and area under the curve (AUC). After applying the search strategy and exclusion criteria, 57 articles are included in the study. The list of variables is long, but mostly non-clinical data are used in the existing literature. A wide range of statistical methods are used, with a recent trend toward machine learning models. The XGBoost model results for the Cox regression in a C-index of 0.87, and the logistic regression on binary LOS has an AUC of 0.94. Many variables identified in the literature are not available at the time of admission, yet they are still used in models for predicting LOS. Machine learning has become the preferred statistical approach in recent studies, though mainly for binary LOS predictions. Based on the current literature, it remains challenging to derive a practical and high performing model for continuous LOS prediction.