Using weak signals to predict spontaneous breathing trial success: a machine learning approach

Area of Science:

• Intensive Care Medicine
• Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Mechanical ventilation weaning is critical in ICU care, with high mortality rates for difficult cases.
• Predicting weaning success is challenging due to complex, multifactorial influences.
• Biosignals, often considered 'weak signals', hold potential for improved prediction.

Purpose of the Study:

• To evaluate machine learning (ML) models for predicting spontaneous breathing trial (SBT) success.
• To identify key biosignals and variables crucial for accurate weaning prediction.
• To explore the application of AI in analyzing complex patient data for better clinical decisions.

Main Methods:

• Retrospective analysis of 232 ICU patients undergoing mechanical ventilation (MV) across two centers.
• Development of ML algorithms using discrete variables and time-series biosignals from the 24 hours preceding SBT.
• Data collected between January 2020 and April 2023.

Main Results:

• Support Vector Classifier (SVC) model demonstrated superior performance in predicting SBT success.
• SVC achieved an AUC-PR of 0.963 and AUROC of 0.922.
• Statistical significance was achieved for both metrics (p < 0.001).

Conclusions:

• ML models utilizing biosignals are effective predictors of SBT success.
• AI offers a promising avenue for developing multidimensional models to analyze weak signals in mechanical ventilation weaning.
• This approach can enhance the prediction of weaning outcomes and potentially reduce mortality.