An Artificial Intelligence Approach to Predict Tracheostomy Requirement in Mechanically Ventilated Critically Ill...

Area of Science:

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

Background:

• Tracheostomy decisions in critically ill patients lack a standardized predictive tool.
• Heterogeneous patient trajectories complicate predicting tracheostomy requirements.
• Prolonged mechanical ventilation and failed weaning are significant concerns.

Purpose of the Study:

• To develop and validate a machine learning model for predicting tracheostomy occurrence in intensive care unit (ICU) patients.
• To identify key clinical parameters influencing early tracheostomy decisions.
• To enhance clinical decision-making for patients requiring mechanical ventilation.

Main Methods:

• Retrospective analysis of 6507 mechanically ventilated ICU patients.
• Development of ten machine learning algorithms using an 80/20 train-test split.
• Performance evaluation using discrimination, calibration, and SHapley Additive exPlanations (SHAP) for interpretability.

Main Results:

• Gradient Boosting model achieved high performance (AUROC 0.92).
• Secretion count emerged as the strongest predictor (14.72%), followed by lactate level, arterial pH, and peak airway pressure.
• SHAP analysis confirmed secretion count, lactate, Glasgow Coma Scale (GCS), and arterial pH as key predictors.

Conclusions:

• Machine learning models integrating early ventilatory and physiological data can meaningfully predict tracheostomy needs.
• The developed model shows potential as an interpretable decision-support tool for prolonged mechanical ventilation.
• Early prediction can aid timely interventions and potentially improve patient outcomes.