Waqar A Sulaiman ¹ Charithea Stylianides ² Andria Nikolaou ^2,3 Zinonas Antoniou ⁴ Ioannis Constantinou ⁴ Lakis Palazis ⁵ Anna Vavlitou ⁵ Theodoros Kyprianou ³ Efthyvoulos Kyriacou ⁶ Antonis Kakas ³ Marios S Pattichis ⁷ Andreas S Panayides ² Constantinos S. Pattichis ^2,3*

• ¹ University of Cyprus, Nicosia, Cyprus
• ² CYENS Centre of Excellence, Nicosia, Cyprus
• ³ Computer Science and Biomedical Engineering Research Centre, University of Cyprus, Nicosia, Cyprus
• ⁴ Research Development Department, 3aHealth, Nicosia, Cyprus, Nicosia, Cyprus
• ⁵ State Health Services Organisation, Nicosia, Cyprus, Nicosia, Cyprus
• ⁶ Cyprus University of Technology, Limassol, Limassol, Cyprus
• ⁷ University of New Mexico, Albuquerque, New Mexico, United States

This study uses machine learning techniques on a large data set in healthcare to meticulously classify the length of stay in the emergency department (ED LOS). Models such as Gradient Boosting (GB), Random Forest (RF), Logistic Regression (LR), and Support Vector Machine (SVM) are employed to categorize ED LOS into short (less than 4.5 hours) or long stays (greater than 4.5 hours) utilizing the MIMIC IV-ED database containing deidentified information of more than 400,000 subjects. GB demonstrated slightly better predictive performance compared to other models, achieving an AUC of 0.736, an accuracy of 70.68%, a sensitivity of 89.50%, and a specificity of 39. 93% on the original data set. For the balanced dataset, GB achieved an AUC of 0.716, an accuracy of 65.11%, sensitivity of 75.05%, and specificity of 57.33%. Emphasizing model interpretability, we employ a novel rule extraction process for the Gradient Boosting model. The rules extracted were based on key predictors of ED LOS such as triage acuity, elixhauser comorbidity index (ECI), arrival methods, and medication history. These extracted rules provide healthcare 1 Sulaiman et al. professionals with clear and actionable information, increasing the transparency of the classification process. By further analyzing the rules, we hope to boost confidence in the model's recommendations, allowing for better decision-making and resource management in emergency departments. Our findings indicate that combining predictive models with rule extraction can significantly improve patient flow and optimize care delivery. However, more work is needed to enhance model performance accompanied by rule interpretability.