Non-Invasive Ventilation Failure in Pediatric ICU: A Machine Learning Driven Prediction.

Area of Science:

• Pediatric Critical Care Medicine
• Biomedical Data Science
• Respiratory Support Technologies

Background:

• Non-invasive ventilation (NIV) is a key strategy to prevent invasive intubation in pediatric intensive care units (PICUs).
• Early identification of NIV failure is critical to mitigate adverse patient outcomes.
• Predicting NIV failure aids in timely clinical decision-making and resource allocation.

Purpose of the Study:

• To identify predictors of first-attempt NIV failure in PICU patients.
• To compare the predictive performance of various machine learning (ML) techniques for NIV failure.
• To develop a reliable ML model for early detection of NIV failure.

Main Methods:

• Utilized data from the TIPNet registry, encompassing patients from 23 Italian PICUs (Jan 2010-Jan 2024).
• Applied and compared ML models: Generalized Linear Models, Random Forest, Extreme Gradient Boosting, and Neural Networks, including an ensemble approach.
• Evaluated model performance using sensitivity, specificity, AUROC, and calibration assessments.

Main Results:

• The Random Forest (RF) model achieved the highest predictive performance with an AUROC of 0.83 and high sensitivity.
• Key predictors for NIV failure included base excess, weight, age, systolic blood pressure, and fraction of inspired oxygen.
• Model calibration confirmed the reliability of predicted NIV failure probabilities.

Conclusions:

• Machine learning models, especially Random Forest, offer highly sensitive prediction of NIV failure in PICU settings.
• The identified predictors provide valuable insights for clinical risk assessment.
• This approach supports timely interventions to improve patient outcomes in pediatric respiratory care.