A BiLSTM-Based Framework for Sleep Spindle and K-complex Detection with Enhanced Segment-Level Learning.

Area of Science:

• Neuroscience
• Signal Processing
• Computational Biology

Background:

• Sleep spindles and K-complexes are crucial EEG signatures for memory, arousal, and neurological health.
• Current detection methods struggle with class imbalance, temporal inconsistencies, and unclear event boundaries.

Purpose of the Study:

• To develop a unified, physiologically informed framework for robust detection of sleep spindles and K-complexes.
• To overcome limitations of existing methods, particularly class imbalance and temporal coherence.

Main Methods:

• Integrated data preprocessing, multimodal feature extraction (spectral and morphological), and temporal modeling using a Bidirectional Long Short-Term Memory (BiLSTM) network.
• Employed context-aware negative filtering and segment-level oversampling to address class imbalance.
• Utilized segment-level dropout and bidirectional hard example mining for improved robustness, with physiologically plausible post-processing for temporal coherence.

Main Results:

• Achieved consistent 3-10% F1-score improvements over state-of-the-art baselines on DREAMS and MASS polysomnography datasets.
• Attained high event-level F1-scores: 0.796 (spindles, DREAMS), 0.938 (K-complexes, DREAMS), 0.882 (spindles, MASS), and 0.903 (K-complexes, MASS).
• Demonstrated superior performance compared to recent deep-learning-based detectors.

Conclusions:

• The proposed framework offers a significant advancement in the accurate and reliable detection of essential sleep EEG events.
• This physiologically informed approach effectively handles class imbalance and enhances temporal consistency in event detection.
• The framework provides a robust tool for neurological assessment and sleep research, outperforming existing methods.