Detection and location of EEG events using deep learning visual inspection

• Published in: PloS one Vol. 19; no. 12; p. e0312763
• Main Author: Fraiwan, Mohammad Amin
• Format: Journal Article
• Language: English
• Published: United States Public Library of Science 23.12.2024; Public Library of Science (PLoS)
• Subjects: Accuracy; Activity patterns; Algorithms; Artificial neural networks; Biology and Life Sciences; Brain - physiology; Classification; Computer and Information Sciences; Datasets; Deep Learning; EEG; Electroencephalography; Electroencephalography - methods; Engineering and Technology; Evaluation; Fourier transforms; Health aspects; Humans; Inspection; Machine learning; Medicine and Health Sciences; Neural networks; Neurological diseases; Object recognition; Performance evaluation; Physical Sciences; Research and Analysis Methods; Sensors; Signal processing; Signal Processing, Computer-Assisted; Sleep; Sleep - physiology; Sleep Stages - physiology; Visual discrimination learning; Waveforms; Wavelet transforms; Jordan
• ISSN: 1932-6203; 1932-6203
• DOI: 10.1371/journal.pone.0312763

The electroencephalogram (EEG) is a major diagnostic tool that provides detailed insight into the electrical activity of the brain. This signal contains a number of distinctive waveform patterns that reflect the subject’s health state in relation to sleep, neurological disorders, memory functions, and more. In this regard, sleep spindles and K-complexes are two major waveform patterns of interest to specialists, who visually inspect the recordings to identify these events. The literature typically follows a traditional approach that examines the time-varying signal to identify features representing the events of interest. Even though most of these methods target individual event types, their reported performance results leave significant room for improvement. The research presented here adopts a novel approach to visually inspect the waveform, similar to how specialists work, to develop a single model that can detect and determine the location of both sleep spindles and K-complexes. The model then produces bounding boxes that accurately delineate the location of these events within the image. Several object detection algorithms (i.e., Faster R-CNN, YOLOv4, and YOLOX) and multiple backbone CNN architectures were evaluated under a wide range of conditions, revealing their true representative performance. The results show exceptional precision (>95% mAP@50) in detecting sleep spindles and K-complexes, albeit with less consistency across backbones and thresholds for the latter.