EEG-based Cross-subject Prediction for Consciousness State Transitions under Sedation using a Deep Learning...

Area of Science:

• Anesthesiology
• Neuroscience
• Artificial Intelligence

Background:

• Intraoperative awareness is a critical concern due to limitations in current anesthesia depth monitoring.
• Traditional binary classifications (conscious/unconscious) fail to capture dynamic transitions during anesthesia induction and emergence.
• Anesthetic transitions vary significantly between individuals and agents.

Purpose of the Study:

• To classify three distinct conscious states: consciousness, transitions, and unresponsiveness.
• To analyze these states during sedation with propofol and midazolam using electroencephalogram (EEG) signals.
• To develop and validate a novel deep learning framework for robust classification.

Main Methods:

• A novel deep learning framework, Deep-ConTrans, was developed.
• The framework incorporates common spatial pattern (CSP) filtering, multi-domain feature extraction, attention-based fusion, and domain-adversarial training.
• Transitions were identified using behavioral responsiveness in patient-controlled sedation paradigms.

Main Results:

• Deep-ConTrans achieved high average classification accuracies: 93.93% (±3.32%) for propofol and 97.42% (±1.68%) for midazolam.
• The model demonstrated strong cross-anesthetic generalizability, maintaining performance across propofol and midazolam.
• EEG signatures of transitions included increased frontal delta and parietal alpha power, consistent with cortical bistability.

Conclusions:

• Deep-ConTrans offers a more granular and reliable method for intraoperative monitoring than binary assessments.
• The model's generalizability and sensitivity enable precise identification of critical transitions, enhancing anesthetic management.
• This approach facilitates personalized sedation protocols based on real-time EEG dynamics, minimizing intraoperative awareness risk.