Predicting sepsis prognosis with machine learning models trained on MIMIC-IV

Abstract

Sepsis remains one of the leading causes of mortality in intensive care units (ICUs), making early prognosis prediction essential for improving clinical outcomes. This study develops and compares five machine learning models—Logistic Regression, Gradient Boosting, XGBoost, LightGBM, and Multi-Layer Perceptron—to predict 30-day mortality in 6,965 sepsis patients from the MIMIC-IV database. Our key contribution is demonstrating that a model-agnostic feature selection approach reduces the variable set from 103 to 20 clinical parameters while maintaining equivalent predictive performance (AUC ¿ 0.87 across all models), significantly improving computational efficiency and clinical interpretability. Systematic hyperparameter optimization with algorithm-specific class-imbalance strategies revealed that ensemble methods (XGBoost, LightGBM, MLP) achieved validation AUC values above 0.90. The identified minimal feature set, dominated by hemodynamic and metabolic markers, provides actionable clinical information while establishing reproducible benchmarks for sepsis mortality prediction, demonstrating the potential of machine learning as a practical support tool for ICU decision-making. Sepsis Mortality Prediction Machine Learning in Healthcare Feature Selection.

Ähnliche Arbeiten