Adding Continuous Vital Sign Information to Static Clinical Data Improves the Prediction of Length of Stay After...

Area of Science:

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

Background:

• Intensive Care Units (ICUs) continuously collect real-time physiologic data from critically ill patients.
• Advancements in big data processing enable the use of this data to predict patient outcomes.
• Predicting specific events during ICU stays can improve patient management.

Purpose of the Study:

• To develop an automated methodology for predicting prolonged mechanical ventilation (>4 days) in pediatric ICU patients.
• To leverage continuous vital sign data from bedside monitors for outcome prediction.
• To assess the performance of machine learning models in this predictive task.

Main Methods:

• Retrospective collection of continuous vital signs and clinical history for 284 pediatric ICU patients.
• Training multiple machine learning models on subsets of patient data.
• Evaluating model performance on unseen validation sets using area under the curve (AUC).

Main Results:

• A model using only vital sign data achieved >83% AUC.
• Combining vital signs with static clinical data improved performance to 90% AUC.
• The developed approach outperformed recent deep learning methods for predicting prolonged ventilation.

Conclusions:

• The proposed workflow offers a scalable approach for real-time predictive systems in ICUs.
• Real-time vital sign data from bedside monitors can be effectively utilized for patient outcome prediction.
• This methodology can aid in designing proactive interventions for critically ill patients.