Predicting Neurological Outcome From Electroencephalogram Dynamics in Comatose Patients After Cardiac Arrest With...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Critical Care Medicine

Background:

• Cardiac arrest survivors often experience coma due to hypoxic-ischemic brain injury.
• Quantitative electroencephalography (EEG) is crucial for prognostic information in these patients.
• Existing EEG analysis methods often miss temporal dynamics.

Purpose of the Study:

• To develop a recurrent deep neural network (RNN) model for predicting long-term neurological outcomes.
• To leverage temporal dynamics from continuous EEG data for improved prognostication.
• To analyze a large international dataset of cardiac arrest patients.

Main Methods:

• Utilized a dataset of continuous EEG recordings from 1,038 cardiac arrest patients.
• Employed a recurrent deep neural network to capture temporal EEG patterns.
• Defined good outcome as Cerebral Performance Category (CPC) 0-2 and poor outcome as CPC 3-5 at 3-6 months.
• Evaluated model performance using 5-fold cross-validation.

Main Results:

• The model's predictive accuracy improved over time, with AUC-ROC reaching 0.88 by 66 hours post-cardiac arrest.
• Achieved high sensitivity and specificity for predicting both good and poor outcomes at 66 hours.
• Demonstrated excellent calibration, with predicted probabilities closely matching observed outcomes (calibration error 0.04).

Conclusions:

• Incorporating EEG temporal trend information significantly enhances the prediction of neurological outcomes.
• The developed RNN model offers a promising tool for prognostication in comatose cardiac arrest patients.
• This approach advances the use of quantitative EEG in critical care neurology.