Systematic review demonstrates deep learning accurately classifies consciousness and brain injury from EEG, highlighting needs for larger datasets and explainable models.
Key Points
To systematically evaluate the state of deep learning applications using EEG signals for assessing consciousness, brain injury, and comatose states.
Conducted a PRISMA-guided systematic review across multiple scientific databases evaluating human studies published between 2018 and 2023.
Screened 475 identified records against defined eligibility criteria, resulting in a final synthesis of 37 included studies.
Most studies targeted classification tasks, centering primarily on evaluating levels of consciousness and detecting epileptic seizures.
Convolutional Neural Networks represented the dominant architecture, while Graph Attention Networks and Transformers emerged as high-performing alternatives.
A majority of investigations evaluated cohorts of fewer than 100 participants, revealing critical gaps in public data availability, multimodal modeling, and explainability.