Visual identification of sleep spindles in EEG waveform images using deep learning object detection (YOLOv4 vs YOLOX)

The electroencephalogram (EEG) is a tool utilized to capture the intricate electrical dynamics within the brain, offering invaluable insights into neural activity. This method is pivotal in identifying potential disruptions in brain cell communication, aiding in the diagnosis of various neurological conditions such as epilepsy and sleep disorders. The examination of EEG waveform morphology and associated characteristics serves as a cornerstone in this diagnostic process. Of particular significance within EEG analysis are sleep spindles, intricate patterns of brain waves implicated in crucial cognitive functions including brain plasticity, learning, memory consolidation, and motor skills. Traditionally, the task of analyzing EEG data has rested upon neurologists, neurosurgeons, or trained medical technicians, a laborious and error-prone endeavor. This study endeavors to revolutionize EEG analysis by leveraging artificial intelligence (AI) methodologies, specifically deep learning object detection techniques, to visually identify and locate sleep spindles within EEG waveform images. The You Only Look Once (YOLOv4) methodology is employed for this purpose. A diverse array of convolutional neural network architectures is meticulously customized, trained, and evaluated to facilitate feature extraction for the YOLOv4 detector. Furthermore, novel YOLOX detection models are introduced and extensively compared against YOLOv4-based counterparts. The results reveal outstanding performance across various metrics, with both YOLOX and YOLOv4 demonstrating exceptional average precision (AP) scores ranging between 98% to 100% at a 50% bounding box overlap threshold. Notably, when scrutinized under higher threshold values, YOLOX emerges as the superior model, exhibiting heightened accuracy in bounding box predictions with an 84% AP score at an 80% overlap threshold, compared to 72.48% AP for YOLOv4. This remarkable performance, particularly at the standard 50% overlap threshold, signifies a significant stride towards meeting the stringent clinical requisites for integrating AI-based solutions into clinical EEG analysis workflows.