Development and external validation of a machine learning model to predict high flow nasal cannula failure.

Area of Science:

• Critical Care Medicine
• Machine Learning in Healthcare
• Respiratory Failure Management

Background:

• High-flow nasal cannula (HFNC) is a vital treatment for acute hypoxic respiratory failure.
• Prolonged HFNC use is linked to adverse outcomes and increased mortality.
• Predicting HFNC failure at initiation remains challenging for clinicians.

Purpose of the Study:

• To develop and externally validate a machine learning model for predicting HFNC failure.
• To compare the predictive discrimination of the ML model against the established ROX Index.

Main Methods:

• A gradient boosting model was trained to predict intubation or death within 24 hours of HFNC initiation.
• The study included over 11,000 adult inpatients across multiple health systems.
• Model performance was evaluated using external validation and compared to the ROX Index.

Main Results:

• The machine learning model achieved an Area Under the Curve (AUC) of 0.760 for predicting 24-hour HFNC failure.
• This performance was significantly superior to the ROX Index (AUC = 0.696, p < 0.001).
• The model's enhanced predictive accuracy was consistent across various time points (2, 6, 12, and 24 hours).

Conclusions:

• A novel machine learning algorithm effectively predicts HFNC failure in patients with acute hypoxic respiratory failure.
• The developed model demonstrates superior predictive performance compared to the ROX Index.
• This algorithm has the potential to enhance clinical decision-making for HFNC therapy.