Evaluating Artificial Intelligence Models for ICU Length of Stay Prediction: A Systematic Review and Meta-Analysis

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Operations Research

Background:

• Intensive care unit (ICU) resource management is vital for balancing patient care and operational efficiency.
• Predicting ICU length of stay (LOS) is crucial for hospital efficiency but challenging with traditional methods.
• Machine learning (ML) and deep learning (DL) offer advanced capabilities for predicting ICU LOS.

Purpose of the Study:

• To systematically review and meta-analyze ML and DL models for predicting adult ICU LOS.
• To assess the predictive performance and identify trends in AI-driven ICU LOS prediction.

Main Methods:

• Systematic review and meta-analysis following PRISMA guidelines.
• Searched eight scientific databases for studies published between 2015 and 2025.
• Included 33 eligible studies on ML/DL for ICU LOS prediction.

Main Results:

• The pooled AUROC was 0.9005, indicating strong predictive capability.
• Most studies focused on binary classification of prolonged ICU stays.
• AI models demonstrated comparable performance in mixed medical-surgical, specialized surgical, and general ICUs.

Conclusions:

• AI-driven models show significant potential for predicting ICU LOS and supporting hospital capacity planning.
• Methodological heterogeneity, lack of external validation, and poor calibration reporting are key limitations.
• Addressing these issues is essential for translating AI predictions into reliable clinical decision support.