A temporal deep learning algorithm for prediction of extubation failures in critical care patients

Area of Science:

• Critical Care Medicine
• Artificial Intelligence in Healthcare
• Machine Learning for Clinical Prediction

Background:

• Extubation failure in intensive care unit (ICU) patients leads to adverse outcomes.
• Current prediction models often lack accuracy due to reliance on static data, failing to capture dynamic disease progression.

Purpose of the Study:

• To introduce TrAcE, a novel deep learning model for enhanced prediction of extubation failure.
• To integrate static and temporal patient data for more accurate and explainable risk assessment.
• To compare TrAcE's performance against traditional methods like the spontaneous breathing test (SBT).

Main Methods:

• Developed a Transformer-based neural network with temporal fusion (TrAcE) for predicting extubation failure.
• Trained and validated the model on the MIMIC-III database and tested on the LOCAL-Ext dataset.
• Assessed model performance using AUROC and AUPRC, with explainability provided by the Captum occlusion method.

Main Results:

• TrAcE achieved high predictive performance with an AUROC of 0.859 (95% CI: 0.836-0.881) on the test set.
• The model demonstrated favorable AUPRC values (0.757 on the test set), indicating robust precision-recall performance.
• TrAcE identified patients suitable for extubation an average of 1.33 days earlier than the SBT in the LOCAL-Ext dataset (p < 0.01).

Conclusions:

• TrAcE offers a powerful, explainable deep learning approach for predicting extubation failure in critical care.
• The model's ability to process dynamic and static data provides real-time risk prediction, improving upon existing methods.
• TrAcE shows potential to optimize extubation timing and improve patient outcomes in the ICU.