Machine Learning Models for Predicting ICU Length of Stay: A Comparative Analysis

Abstract

Proper estimation of intensive care unit (ICU) length of stay (LOS) is important for effective resource utilization, patient flow management, and clinical decision assistance. In this study, a comparative analysis of machine learning models: Linear Regression, Random Forest, XGBoost, and Multilayer Perceptron (MLP) were conducted to assess the prediction efficiency of these models on the publicly available MIMIC-IV dataset. The information, demographic, clinical, and lab features, were pre-processed and divided into training, validation, and test sets (70%, 15%, and 15%). Performance of the model was measured in terms of various regression metrics including mean absolute error (MAE), root mean squared error (RMSE), coefficient of determination (R²), mean absolute percentage error (MAPE), and median absolute error (MedAE). Results showed that the best in prediction were ensemble-based models, namely XGBoost (MAE = 2.14 days, RMSE = 2.95, R² = 0.61) over Random Forest and MLP, and the worst was Linear Regression with no ability to detect nonlinear relationships. This research emphasizes the clinical value of achieving an average error of less than two days, which allows hospitals to predict ICU demand and enhance staffing and resource allocation. Additionally, the importance of interpretability was mentioned, particularly regarding the necessity for transparent models that foster clinician trust and encourage their adoption. Overall, the findings point to the promise of future machine learning systems in designing better ICU resource management and also suggest that greater emphasis needs to be given to transparency and generalizability in the application to clinical environments.