Interdisciplinary seminar by Alex Groh

Speaker: Alex Groh (University of Aberdeen)

Title: A data driven approach to hospital length of stay prediction and capacity modelling

Abstract: The uncertainty about the volume and the length of stay (LOS) of non-elective admissions poses a substantial barrier to efficient hospital organisation and planning. Predictive modelling of LOS using electronic health records can support resource management, clinical care, and patient flow optimisation. In this talk I will discuss the utility of machine learning (ML) models for reliable LOS prediction at Aberdeen Royal Infirmary (ARI), the largest hospital in Aberdeenshire. We analysed NHS Grampian hospital data from Aberdeen Royal Infirmary, comprising 189,620 non-elective inpatient admissions between 2015-2020 and we evaluated random forest (RF) and eXtreme Gradient Boosting (XGBoost) algorithms on both regression and multi-class classification prediction problems for general hospital population and within individual medical specialties. These models were developed using variables available within 24 hours of admission and I will additionally discuss the effect of diagnostic features on predictive performance of our models. Specialty-based stratification revealed large differences in predictive performance across medical specialties in both regression and classification tasks. In particular, the models achieved strong performance in Orthopaedics but performed poorly in Geriatric Medicine. Overall, our results suggest that variables available on the first day of admission are insufficient for accurate prediction of continuous LOS, but we will show that a coarse-grained classification model offers a significant enhancement in predictive performance to support clinical and operational workflows in healthcare facilities.