MACHINE LEARNING MODELS FOR PREDICTING PATIENT OUTCOMES IN INTENSIVE CARE UNITS (ICU): A CLINICAL APPROACH

Abstract

The integration of machine learning (ML) models into Intensive Care Units (ICUs) offers transformative potential for predicting critical patient outcomes and enhancing clinical decision-making. This study developed and evaluated five ML algorithms—Logistic Regression, Random Forest, XGBoost, Support Vector Machine (SVM), and Neural Networks—using a dataset of 25,000 ICU patients from the MIMIC-IV database. The goal was to predict mortality, length of stay (LOS), and 30-day readmission risk. Among all models, XGBoost demonstrated superior performance with an accuracy of 0.87, F1-score of 0.82, and AUC of 0.91 for mortality prediction. SHAP (SHapley Additive exPlanations) analysis identified SOFA score, age, and lactate level as the most influential predictors.For both readmission prediction (AUC = 0.87  >>Calibration tests showed the forecast estimates were precise; Brier score of the XGBoost model was 0.139 and its slope was nearly the same as 1.0.  The use of ROC curves, confusion matrices, calibration plots, and SHAP summaries, as well as other graphical representations, helped explain the model, which made it easier to interpret how predictions were developed.  The adoption of transparent AI approaches increased comprehension of model logic, making it more useful in clinical rooms and contributing to practitioners’ trust.  This study shows that interpretable machine learning models, such as XGBoost, can provide accurate, reliable and actionable predictions that may assist with timely interventions and optimal resource distribution in intensive care units.  The findings provide evidence of the capability to embed these interpretable models into realtime ICU workflow and pave the way for future development to federated learning and multi-modal systems.