YOLOv11-Based Detection of Stage N2 Sleep-Specific Neurophysiological Events

• Published in: Biomedical Engineering International Conference pp. 1 - 5
• Main Authors: Ryadi, Kezia Sarahsophia Immanuel, Lin, Che-Wei
• Format: Conference Proceeding
• Language: English
• Published: IEEE 15.07.2025
• Subjects: Accuracy; Biomedical Deep Learning; Brain modeling; Classification algorithms; Continuous wavelet transforms; EEG Signal Processing; Electroencephalography; Signal processing algorithms; Sleep; Sleep Spindle and K-Complex Detection; Sleep Stage Classification; Testing; Training; Wavelet analysis; YOLOv11
• ISSN: 2473-7607
• DOI: 10.1109/BMEiCON66226.2025.11113667

Sleep disturbances are a growing global health concern, with a particularly high prevalence in Taiwan. Among various sleep stages, non-rapid eye movement stage 2 (N2) is strongly associated with sleep quality and is characterized by specific neurophysiological events such as sleep spindles and K-complexes. Accurate detection of these events is essential for sleep stage classification and the diagnosis of sleep-related disorders. This study proposes a YOLOv11-based object detection algorithm for automatic identification of sleep-specific events in electroencephalogram (EEG) signals. EEG recordings were sourced from the DREAMS database, with sleep spindle and K-complex annotations verified by clinical experts. Signals from the CZA1 channel were segmented into 5-second windows, and class imbalance was addressed using random oversampling, resulting in 3,480 spindle and 12,858 K-complex training samples. The algorithm integrates Tunable Q-factor Wavelet Transform (TQWT), Morphological Component Analysis (MCA), and Continuous Wavelet Transform (CWT) for effective feature decomposition and classification. Model performance was evaluated using a Leave-One-Subject-Out (LOSO) cross-validation strategy, achieving F1-scores of 0.902 for sleep spindle and 0.978 for K-complex detection. Additional testing on an AI-generated mixed-event sequence confirmed the model's robustness, yielding F1-scores of 0.844 and 0.939, respectively. The proposed system, with a compact model size of 18.3 MB, demonstrates high accuracy and efficiency, providing a promising solution for real-time sleep monitoring and automated sleep staging.