tsta

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Multiclass ICU Length-of-Stay Prediction Using Tree-Based Machine Learning Techniques

Authors :

Mahyar Mohammadian¹ Somayeh Afrasiabi²

1- School of Electrical and Computer Engineering, Shiraz University 2- School of Electrical and Computer Engineering, Shiraz University

Keywords :

multi-class prediction،ICU length of stay،CatBoost،MIMIC III،Area Under Curve

Abstract :

Accurate prediction of intensive care unit (ICU) length-of-stay (LOS) is essential for patient management and resource planning. This study compares four tree-based machine learning models—Random Forest, XGBoost, LightGBM, and CatBoost—for multiclass LOS prediction using the MIMIC-III database. A total of 42,306 ICU stays were processed with 17 physiologic variables and discretized into 10 ordered LOS classes. Models were evaluated using quadratic-weighted Cohen’s kappa (κ) and Mean Absolute Deviation (MAD) to capture ordinal agreement and temporal accuracy. CatBoost achieved the best performance (κ = 0.444, MAD = 124.66 hours), effectively predicting both short- and longstay patients, which are operationally critical. XGBoost and Random Forest provided intermediate results, while LightGBM showed lower temporal precision (MAD = 164.19 hours). The results demonstrate that CatBoost’s ordered boosting strategy and native handling of categorical variables enable robust, interpretable predictions suitable for clinical and operational decision-making. These findings highlight the potential of tree-based machine learning to transform ICU LOS prediction from a retrospective metric into a proactive, reliable and interpretable tool for optimizing patient flow, resource allocation and decision-making. The study provides a foundation for future improvements using richer time-series data, multimodal inputs, and multicenter validation.