A continuous convolutional neural network: very early EEG most predictive for poor neurological outcome in postanoxic...

Area of Science:

• Neuroscience
• Artificial Intelligence in Medicine
• Critical Care Medicine

Background:

• Accurate prognostication of neurological outcome after cardiac arrest (CA) is crucial for patient management.
• Electroencephalography (EEG) provides valuable insights into brain function post-CA.
• Convolutional Neural Networks (CNNs) show promise in analyzing complex biological signals like EEG.

Purpose of the Study:

• To evaluate the predictive capability of a continuous CNN model using EEG data over time after CA.
• To compare the performance of 9-electrode versus 4-electrode EEG configurations for predicting poor neurological outcome.

Main Methods:

• A CNN model was trained on hourly EEG epochs from 366 postanoxic coma patients up to 72 hours post-CA.
• Neurological outcome was assessed using the Cerebral Performance Category (CPC) scale within 6 months.
• Model performance was evaluated based on AUC and sensitivity at 0% FPR, comparing cumulative and independent predictions.

Main Results:

• The 9-electrode cumulative CNN model achieved an AUC of 82% and 71% sensitivity at 0% FPR.
• Optimal prediction occurred within 8 hours post-CA (AUC=0.93, sensitivity=0.81 at 0% FPR).
• No significant performance difference was found between 9-electrode and 4-electrode configurations.

Conclusions:

• CNNs are valuable tools for predicting poor neurological outcome in postanoxic coma patients, particularly with early EEG data (6-14 hours post-CA).
• A reduced 4-electrode EEG setup is as effective as a full 9-electrode setup for prognostication.
• Further research can refine CNNs by incorporating temporal EEG signal dependencies.