Transfer learning approach for sleep stage classification with limited training data

Adapting sleep stage classification models to new subjects typically requires extensive labeled data. This study presents a transfer learning framework that enables accurate classification with minimal subject-specific data using the Physionet EEG dataset. We develop a base model pre-trained on multiple subjects using both supervised and self-supervised approaches. Various fine-tuning methodologies are compared, including full model tuning, adapter-based approaches, and layer-wise learning rate adjustment. Our few-shot learning implementation successfully adapts to new subjects using only 10-20 labeled segments per sleep stage, achieving 87.3% accuracy compared to 91.8% with full data fine-tuning. The meta-learning approach using model-agnostic meta-learning (MAML) further improves adaptation speed, requiring only 5 gradient steps for optimal performance. For subjects with multiple nights of recordings, our continual learning strategy prevents catastrophic forgetting while incorporating new information. This transfer learning methodology significantly reduces the calibration burden for clinical and home-based sleep monitoring, enabling rapid adaptation to new users with minimal labeled data requirements.