Improving Patient Trajectory Forecasts in Hospitals: Using Emergency Department Data for Length of Stay Prediction and Next Hospital Unit Classification

Accurately forecasting a patient’s trajectory during hospitalization is essential for effective hospital management. Predicting length of stay (LOS) and next hospital unit can assist in resource planning and management, benefiting patients, physicians, and hospitals. This paper extends previous research on LOS prediction and introduces the task of next hospital unit classification. The study utilizes the MIMIC dataset, which now includes specific emergency department (ED) data, making it suitable for machine learning methodologies. Several contributions are made, including the addition of the next hospital unit classification task, an extended related work section, an expanded dataset description, and a thorough error analysis. The CatBoost model is employed to handle the high-dimensional categorical features in the dataset, along with feature engineering, hyperparameter tuning, and a customized loss function for the LOS regression task. Benchmarking against baseline models and related research demonstrates the superior performance of the proposed methods, with an average absolute error of 2.36 days for LOS prediction and a 50% accuracy for the next hospital unit classification. The paper provides a comprehensive overview of the related work, dataset description, approach, results, and concludes with insights and future directions.