Benchmarking Predictive Models in Electronic Health Records: Sepsis Length of Stay Prediction

Abstract

Forecasting Sepsis length of stay is a challenge for hospitals worldwide. Although there are many attempts to improve sepsis length of stay prediction; however, there is still lack of baselining prediction metrics that can give better results for sepsis length of stay prediction in management hospital systems. This paper introduces a research architecture to predict and benchmark the Length of Stay (LOS) for Sepsis diagnoses from electronic medical records using the machine learning models. The architecture considered the time factor to identify the outperforming algorithms for Sepsis LOS prediction. This work contributes to the field of predictive modelling and information visualization for hospital management systems. Our results showed that the ensemble methods in particular the random forest (RF) outdo other classification models to predict the LOS for Sepsis from electronic medical records for Intensive Care Unit “ICU”-based hospitalizations.