Health outcome predictive modelling in intensive care units

Abstract

The literature in Intensive Care Units (ICUs) data analysis focuses on predictions of length-of-stay (LOS) and mortality based on patient acuity scores such as Acute Physiology and Chronic Health Evaluation (APACHE), Sequential Organ Failure Assessment (SOFA), to name a few. Unlike ICUs in other areas around the world, ICUs in Ontario, Canada, collect two primary intensive care scoring scales, a therapeutic acuity score called the “Multiple Organs Dysfunctional Score” (MODS) and a nursing workload score called the “Nine Equivalents Nursing Manpower Use Score” (NEMS). The dataset analyzed in this study contains patients’ NEMS and MODS scores measured upon patient admission into the ICU and other characteristics commonly found in the literature. Data were collected between January 1st, 2015 and May 31st, 2021, at two teaching hospital ICUs in Ontario, Canada. In this work, we developed logistic regression, random forests (RF) and neural networks (NN) models for mortality (discharged or deceased) and LOS (short or long stay) predictions. Considering the effect of mortality outcome on LOS, we also combined mortality and LOS to create a new categorical health outcome called LMClass (short stay & discharged, short stay & deceased, or long stay without specifying mortality outcomes), and then applied multinomial regression, RF and NN for its prediction. Among the models evaluated, logistic regression for mortality prediction results in the highest area under the curve (AUC) of 0.795 and also for LMClass prediction the highest accuracy of 0.630. In contrast, in LOS prediction, RF outperforms the other methods with the highest AUC of 0.689. This study also demonstrates that MODS and NEMS, as well as their components measured upon patient arrival, significantly contribute to health outcome prediction in ICUs.