Abstract
This study aims to predict the mortality rate among septic patients with early-onset hypoalbuminemia in the intensive care unit (ICU) using machine learning algorithms. Utilizing patient data from the MIMIC-IV and eICU databases, we divided MIMIC-IV samples into training and internal validation sets, with eICU samples serving as an external validation set. We developed the predictive model using various feature selection techniques and machine learning algorithms, and evaluated its performance using metrics such as AUC, accuracy, precision, recall, and F1 score. The SHAP method was used for model interpretability. The CatBoost model, developed using recursive feature elimination, outperformed other algorithms, demonstrating robust generalization with AUC values of 0.845, 0.746, and 0.827 across the respective datasets. This pioneering study presents a machine learning model with high accuracy and robust extrapolation capabilities for predicting mortality rates in septic patients with early-onset hypoalbuminemia in the ICU, providing valuable decision support for clinicians.