Factors associated with length of stay following an emergency medical admission

•Factors predicting LOS were examined for all emergency medical admissions.•Predictors of LOS were identified; a generalizable truncated LOS model was presented.•More complicated patient outcome predictors significantly associated with LOS.•Reducing LOS would be challenging given the ageing population.•Predicting LOS can offer a starting point for identifying efficiencies.

BackgroundHospitals are under pressure to use resources in the most efficient manner. We have examined the factors predicting Length of Stay (LOS) in one institution, using a database of all episodes of emergency medical admissions prospectively collected over 12 years.AimTo examine the ability to predict hospital LOS following an emergency medical hospital admission.MethodsAll emergency admissions (66,933 episodes; 36,271 patients) to St. James's Hospital, Dublin, Ireland over a 12-year period (2002–2013) were evaluated in relation to LOS. Predictor variables (identified univariately) were entered into a multiple logistic regression model to predict a longer or shorter LOS (bivariate at the median). The data was also modelled as count data (absolute LOS), using zero truncated Poisson regression methodology. Appropriate post-estimation techniques for model fit were then applied to assess the resulting model.ResultsThe major predictors of LOS included Acute Illness Severity (biochemical laboratory score at admission), Charlson co-morbidity, Manchester Triage Category at admission, Diagnosis Related Group, sepsis status (based on blood culture result), and Chronic Disease Score Indicator. The full model to predict a LOS above or below the median had an Area Under Receiver Operating Characteristic (AUROC) of 0.71 (95% CI: 0.70, 0.71). The truncated Poisson model appeared to achieve a good model fit (R2 statistic = 0.76).ConclusionPredictor variables strongly correlated with LOS; there were linear increases within categories and summation between variables. More predictor variables may improve model reliability but predicting LOS ranges or quantiles may be more realistic, based on these results.