Performance Comparison of Models for Short-Term Mortality Prediction in ICU Patients

Early prediction of mortality risk in intensive care unit (ICU) patients is critical for clinical decision-making, but existing prediction models still have room for improvement in analyzing short-term physiological indicators.This study, based on the MIMIC-IV Demo dataset, examines the performance differences of machine learning and deep learning models in short-term mortality prediction. Dynamic physiological indicators and static clinical data from the first 24 hours of admission were used to compare the predictive performance of XGBoost, Random Forest, and LSTM models for 7-day mortality in ICU patients. The results showed that XGBoost significantly outperformed other models in key metrics such as precision, AUC, and recall. However, LSTM's prediction performance was relatively weak due to limited data and a short observation window. The study suggests that XGBoost is more suitable for predicting short-term ICU mortality risk in current small-scale clinical data. This research provides a reference for the development of real-time clinical early warning systems. It also suggests that further improvements are needed in the future, including larger sample sizes and more optimized time series modeling methods, to enhance prediction accuracy.