Collaborative Reasoning Framework for Edge-Deployable EEG Sleep Staging via Local LLM
Cheng-Lin Cheng, T.Y. Lin, Chia-Kai Chang · 2025
Manual EEG-based sleep staging, crucial for sleep medicine, is time-consuming and variable. This work focuses on an efficient automated approach using deep learning. We employ Continuous Wavelet Transform (CWT) to generate spectrograms from single-channel EEG data, avoiding complex manual feature engineering. These spectrograms are classified by fine-tuning a pretrained Vision Transformer Hybrid (ViTHybrid) model, achieving up to$\mathbf{4} \times$faster training and$\sim \mathbf{1 0 \%}$higher performance than training from scratch. Validated on a clinical dataset of$\sim 30,000$epochs with high signal variability, the single-channel fine-tuned model achieves a highest 66.7 % Top-1 and over 93% Top-2 accuracy. In addition to singlechannel operation, we further integrate multi-channel classification results using a locally hosted Large Language Model (LLM) for decision-level reasoning, enhancing robustness while maintaining full offline data privacy. The open-source based, fine-tuned model provides a robust, efficient, and privacy-aware foundation for automated sleep staging.