Approaching machine learning for length of stay prediction in sepsis patients presenting to emergency department from 2014-2024

Sepsis is a life-threatening presentation in the emergency department and associated with poor clinical outcomes and emergency crowding by prolonged length of stay. The current study outlines the length of stay prediction in sepsis patients using machine learning methods with a decade long analysis from 2014-2024. The study deployed four different classification algorithms for prediction of length of stay: Random Forest, Logistic Regression, Gradient boosting, and XG Boost. All the models were evaluated through Sensitivity, Specificity, Precision, F1 score and Area under curve. The dataset was randomly split into a training set (70%) and a validation set (30%). A total of 13,858 adult patients (>18 years) were included in the study. A total of 14 key attributes were analyzed. No difference in the length of stay was observed in the patients with age >60 or <60 years. Interestingly there was equal number of patients with sepsis in younger and older age groups. The Random Forest classifier best predicted both short and long LOS (Sensitivity: 75.54% and Specificity: 78.35%, AUC: 94.5%). The external validation of the model yielded an overall accuracy of 80% and sensitivity and specificity of 68% and 91%, respectively. The model can be implemented for improving ED workflows and operational efficiency. Future studies with larger and more diverse datasets can further improve the model performance.