Predicting length of stay, discharge destination and mortality of patients with hip fractures

Abstract

Background. In the Netherlands, the annual incidence of hip fractures is 88 per 100,000. A major issue in healthcare concerns the shortage of beds in hospitals, caused by a decreased flow to nursing homes. Consequently, patients have a longer hospital stay, which could lead to unnecessary complications and a longer rehabilitation period. Therefore, optimization of the patient flow is desired. Aim. In this study, we investigated which variables are relevant for the prediction of length of stay (LOS), discharge destination and mortality. Moreover, we investigated different models on their predictive performance. Methods. Various methods have been applied to achieve the goal, namely: literature study, interviews, model development and statistical analysis (ANOVA). We compared regression, lasso regression and random forest (RF) models with and without feature selection. Results. This research showed that age, fracture type and involvement of geriatrician are important predictors for LOS. The most suitable model was RF without feature selection. Furthermore, it showed that age, involvement of geriatrician and living situation prior to the injury are important predictors for discharge destination. The best model was RF without feature selection. Next, it showed that age, dementia and pre-surgery mobility are important predictors for mortality. Lastly, statistical tests showed that the best models were not significantly better than all other models included in the comparison. Conclusion. These findings suggest that RF without feature selection could be used in patient flow optimization for hip fracture patients. However, these are not statistically significant and therefore the models could be improved.