Predicting Prolonged Length of Stay for Hospitalized Stroke Patients Using Common Physiological Features

Background and PurposeStroke is a leading cause of death and disability worldwide. Predicting which patients are at risk for a prolonged length of stay (LOS) could assist in coordination of care and serve as a rough measure of clinical recovery trajectory. During the acute stroke period, there is a disruption in the fidelity of the blood-brain barrier and cerebral autoregulation, and we hypothesize that trends in physiologic parameters early in a patients hospital course may be used to predict which patients are increased risk for a prolonged LOS. In this work we sought to create a model to predict prolonged LOS (defined as [≥] 7 days) from patient data available at admission as well as routinely collected physiologic (pulse, blood pressure, respiratory rate, temperature), and other data from the first 24 hours of admission. MethodsThis retrospective cohort study included stroke patients admitted to an urban comprehensive stroke center between 2016-2019. Data included common physiological parameters (pulse, temperature, blood pressure, respirations, and oxygen saturation) as well as demographic and comorbidity data. Raw time series data were transformed into statistical features for modeling. Logistic regression, random forest, and XGBoost models were trained on data collected during the first 24 hours after hospital admission to predict prolonged LOS and evaluated on a held-out test set. ResultsA total of 2,025 patients were included. Using an XGBoost classifier we obtained a ROC AUC of 0.85 and Precision-Recall AUC of 0.77, with the optimal operating point achieving an accuracy of 0.80, sensitivity of 0.78, specificity of 0.81. ConclusionsThe model suggests that prolonged LOS can be predicted with reasonable accuracy using clinical data obtained within the first 24 hours of hospitalization. This approach could provide the basis for development of a risk score and augment the care coordination process.