Efficacy prediction of noninvasive ventilation failure based on the stacking ensemble algorithm and autoencoder.

Area of Science:

• Critical Care Medicine
• Biomedical Data Science
• Machine Learning in Healthcare

Background:

• Accurate early prediction of noninvasive ventilation (NIV) failure is crucial for timely treatment adjustments in intensive care unit (ICU) patients.
• Challenges in predicting NIV efficacy stem from time-series correlated clinical data and imbalanced class distributions.
• Existing methods struggle with precise and early prediction of NIV failure, especially in severe patient populations.

Purpose of the Study:

• To develop a precise model for predicting the probability of noninvasive ventilation failure.
• To enable early prediction of NIV failure within the initial 1-2 hours of therapy.
• To identify key clinical features influencing NIV failure prediction and understand their correlations.

Main Methods:

• Proposed a Stacking and Modified SMOTE algorithm for Noninvasive ventilation failure prediction (SMSN) model.
• Utilized a Long Short-Term Memory (LSTM) autoencoder for automatic extraction of time-series features.
• Employed a modified SMOTE algorithm to handle imbalanced classes, combined with logistic regression, random forests, and Catboost classifiers via stacking ensemble.

Main Results:

• The SMSN model was trained on data from 2495 patients, with 1996 in the training set and 499 in the testing set.
• Achieved a high prediction performance with an F1 score of 79.4% and an accuracy of 88.2%.
• Demonstrated significant improvements over traditional logistic regression, with a 4.7% increase in F1 score and a 1.3% increase in accuracy.

Conclusions:

• The SMSN model provides accurate and early prediction of noninvasive ventilation failure.
• SHAP analysis identified oxygenation index, pH, and FiO2 after 1 hour as critical predictors of NIV failure.
• The findings support the clinical utility of the SMSN model for optimizing critical care management of patients on NIV.