Predicting Prolonged Length of ICU Stay through Machine Learning

Area of Science:

• Critical Care Medicine
• Health Informatics
• Machine Learning in Healthcare

Background:

• Predicting prolonged length of stay (pLOS) in intensive care units (ICU) is crucial for resource allocation and patient management.
• Existing scoring systems may not fully capture the complexity of patient data for accurate pLOS prediction.

Purpose of the Study:

• To develop and validate machine learning (ML) models for predicting pLOS in general ICU patients.
• To compare the performance of different ML models against a traditional scoring system.

Main Methods:

• Utilized the eICU and MIMIC III databases for internal and external validation, respectively.
• Developed four ML models: random forest, support vector machine, deep learning, and gradient boosting decision tree (GBDT).
• Evaluated models using AUROC, AUPRC, ECI, and Brier score, comparing them against the SAPS II score.

Main Results:

• The GBDT model demonstrated superior performance in both internal and external validation, showing excellent discrimination and calibration.
• All four ML models outperformed the customized SAPS II model in predicting pLOS.
• The GBDT model achieved a Brier score of 0.164 (internal) and 0.166 (external), with AUROCs of 0.742 and 0.747, respectively.

Conclusions:

• The GBDT-based pLOS-ICU prediction model is highly effective and reliable for identifying at-risk patients.
• This ML model has the potential to support clinical decision-making and improve patient outcomes through timely interventions.