Detection and location of EEG events using deep learning visual inspection

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Electroencephalogram (EEG) is crucial for understanding brain activity, with specific waveforms like sleep spindles and K-complexes indicating neurological states.
• Current methods for identifying these EEG events are often specialized and have performance limitations.

Purpose of the Study:

• To develop a unified AI model capable of detecting and localizing both sleep spindles and K-complexes simultaneously.
• To mimic the visual inspection approach used by specialists for event identification.

Main Methods:

• Evaluation of object detection algorithms (Faster R-CNN, YOLOv4, YOLOX) and various Convolutional Neural Network (CNN) backbones.
• Training a single model to generate bounding boxes for precise event localization within EEG data.

Main Results:

• Achieved exceptional precision (>95% mAP@50) in detecting both sleep spindles and K-complexes.
• Demonstrated the model's ability to delineate event locations accurately using bounding boxes.

Conclusions:

• The novel approach offers a significant improvement over traditional methods for identifying key EEG waveform patterns.
• The developed model provides a robust tool for sleep studies and neurological disorder diagnostics.