Pediatric Intensive Care Unit Length of Stay Prediction by Machine Learning.

Area of Science:

• Pediatric Intensive Care
• Machine Learning
• Health Informatics

Background:

• Predicting length of stay (LOS) in the Pediatric Intensive Care Unit (PICU) is crucial for resource and staffing management.
• Traditional statistical methods and human predictions have limitations in accuracy for LOS prediction.

Purpose of the Study:

• To develop and validate machine learning (ML) models for predicting PICU LOS.
• To compare the accuracy of various ML algorithms against traditional methods.

Main Methods:

• A retrospective analysis of 123,354 patient encounters from the Virtual Pediatric Systems (VPS) database (2015-2020).
• Evaluation of ML models including Gradient Boosting, CatBoost, and Recurrent Neural Networks (RNNs) for LOS prediction at various time thresholds.
• Comparison of ML model performance against traditional prediction methods.

Main Results:

• Gradient Boosting, CatBoost, and RNN models achieved the highest accuracy (70-73%) in predicting PICU LOS, particularly at 36h and 48h thresholds.
• These ML models significantly outperformed traditional methods, which reported around 50% accuracy.
• The models demonstrated balanced sensitivity (up to 74%) and specificity (up to 82%).

Conclusions:

• Machine learning models, especially Gradient Boosting, CatBoost, and RNNs, show moderate effectiveness in predicting PICU LOS.
• These models offer a significant improvement over existing prediction methods and human estimations.
• Further refinement with specialized databases may enhance clinical applicability and accuracy for PICU management.