An Artificial Intelligence Approach to Predict Tracheostomy Requirement in Mechanically Ventilated Critically Ill Patients: A Retrospective Single-Center Study

TL;DR

This study uses AI to predict whether critically ill patients on ventilators will need a tracheostomy, helping doctors make earlier decisions.

Contribution

The novel contribution is a machine learning model that accurately and interpretably predicts tracheostomy need using early ventilatory and physiological data.

Findings

Gradient Boosting achieved high accuracy (AUROC 0.92) in predicting tracheostomy requirements.

Secretion count was the strongest predictor, followed by lactate level and arterial pH.

SHAP analysis provided interpretable insights into how individual features influence predictions.

Abstract

Background: In critically ill patients, tracheostomy decisions are driven by heterogeneous and dynamic clinical trajectories, and no universally accepted scoring system exists to reliably predict tracheostomy requirement. An accurate and interpretable prediction model could help earlier decision-making and potentially reduce prolonged mechanical ventilation (MV) and failed weaning. Methods: In this retrospective study, data from 6507 mechanically ventilated intensive care unit (ICU) patients were analyzed using an electronic clinical decision support system; 1049 patients required tracheostomy and 5458 did not. The primary outcome was the prediction of tracheostomy occurrence during ICU stay based on invasive mechanical ventilation (IMV) parameters obtained within the first five days. The secondary outcome was the identification of the most influential parameters guiding tracheostomy decision-making during early IMV. Ten machine learning algorithms were developed using an 80/20 train–test split. Model performance was assessed using discrimination, calibration, and clinical performance metrics. Explainability was evaluated using SHapley Additive exPlanations (SHAP) analysis. Results: Among all models, Gradient Boosting demonstrated strong discrimination and calibration performance (AUROC 0.92, AUPRC 0.56, specificity 97%, F1 score 0.46, Brier score 0.078). In the Gradient Boosting model, feature importance analysis demonstrated that secretion count was the strongest predictor of tracheostomy requirement, accounting for 14.72% of the model’s predictive contribution. This was followed by lactate level (6.12%), arterial pH (3.74%), and peak airway pressure (3.57%). SHAP-based analyses consistently identified secretion count as the strongest predictor of tracheostomy requirement, followed by lactate level, Glasgow Coma Scale (GCS), and arterial pH. In addition, SHAP provided clinically interpretable insights into the direction and magnitude of the effects of individual predictors. Conclusions: Machine learning models integrating early-phase ventilatory and physiological data may enable clinically meaningful prediction of tracheostomy requirement. The combination of strong performance and explainability suggests potential utility as a decision-support tool in critically ill patients requiring prolonged MV.

Figures