A Predictive Framework for ICU Stay Classification using EHR and Explainable AI Techniques

Abstract

Effective bed management in hospitals reduces expenses and enhances patient outcomes. This paper presents a prediction model of the LOS in ICU at admission based on EHR data. The experiment evaluates numerous ML techniques, such as the LR, the Random ForRFest, the MLP, the Gradient Boosting, the XGBoost, and an extension based on the CatBoost, which are used in the hospital stay dataset of the Kaggle repository. Evaluations of the algorithms use AUC, accuracy, precision, recall, and F1-score. XGBoost was the most accurate of the traditional algorithms, but the improved CatBoost approach was superior to all with an accuracy of 98.25. XAI methods such as SHAP were used in order to explain feature contributions. The study demonstrates how patient EHR data and advanced ML models can be used to predict ICU admissions to enable improved allocation of resources in healthcare organizations.