Predicting mechanical ventilation duration in ICU patients: A data-driven machine learning approach for clinical...

Area of Science:

• Critical Care Medicine
• Artificial Intelligence
• Health Informatics

Background:

• Mechanical ventilation is crucial in ICUs but associated with complications and high costs.
• Predicting the duration of mechanical ventilation using clinical data is challenging.
• Accurate prediction aids in clinical decision-making, resource management, and planning procedures like tracheostomy.

Purpose of the Study:

• To develop explainable artificial intelligence (AI) models for predicting mechanical ventilation duration.
• To leverage diverse clinical parameters from ICU patient data for prediction.
• To enhance the interpretability of AI models in critical care settings.

Main Methods:

• Development and testing of regression and classification AI models using data from 323 mechanically ventilated patients.
• Feature selection was performed using Shapley Additive Explanations (SHAP) on a random forest model.
• Models were trained and validated using 5-fold cross-validation.

Main Results:

• A least-squares boosting regression model predicted ventilation duration with an R² of 0.65, identifying tracheostomy as a key predictor.
• A K-nearest neighbors classification model achieved 79.1% accuracy and an AUROC of 0.82 for predicting short-term vs. long-term ventilation.
• Important predictors included ICU admission type, PO2, and pH.

Conclusions:

• AI-driven prediction of ventilation duration can optimize ICU workflows and resource utilization.
• Explainable AI, particularly SHAP-based feature selection, can facilitate clinical adoption.
• These models offer potential for improved personalized patient care in ICUs.