The Prognostic Performance of Artificial Intelligence and Machine Learning Models for Mortality Prediction in...

Area of Science:

• Critical care medicine
• Medical informatics
• Data science

Background:

• In-hospital mortality prediction in intensive care units (ICUs) is a significant clinical challenge.
• Existing clinical scoring systems have limitations in accuracy and timeliness.

Purpose of the Study:

• To systematically review the application and performance of artificial intelligence (AI) and machine learning (ML) models for predicting in-hospital mortality in ICU patients.
• To compare the efficacy of AI/ML models against traditional clinical scoring systems.

Main Methods:

• Systematic review following PRISMA guidelines, analyzing 15 studies (Jan 2015-Apr 2025).
• Focus on AI/ML algorithms (XGBoost, random forest, logistic regression, deep learning) using MIMIC and eICU-CRD databases.
• Evaluation of predictive features and model performance.

Main Results:

• AI/ML models consistently outperformed traditional systems (APACHE, SOFA, SAPS) in mortality prediction.
• Ensemble methods (XGBoost, random forest) and deep learning showed high accuracy.
• Key predictors included age, vital signs, lab values, and neurological status.
• Models using data from the first 24 hours demonstrated feasibility for early risk stratification.

Conclusions:

• AI and ML models possess significant clinical potential for enhancing ICU mortality prediction.
• Future research should focus on prospective validation, standardization, and clinical implementation.
• These technologies can potentially improve decision-making and patient outcomes in critical care.