NarratingSleep: A Language Model-Based Framework for Sleep Staging

Automatic sleep staging is significant for accurately diagnosing sleep disorders, improving sleep quality, and advancing research on sleep patterns. Current methodologies for automatic sleep staging require substantial quantities of labeled data for training models from scratch. In this study, we proposed a framework called NarratingSleep that utilizes the pre-trained language model’s ability to process sequential data for automatic sleep staging. NarratingSleep employs a cross-modal transfer learning approach and does not require altering the structure of the language model, enabling training by using various language models. We bridged the gap between continuous time series measurements and discrete word tokens by implementing a time series vocabulary V<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ts</inf>. Additionally, we employed a temporal content classifier for many-to-many classification. By learning the temporal context of the current frame, we significantly improved accuracy. After conducting several experiments, we show that the NarratingSleep: (a) achieves SOTA performance across multiple datasets using fewer training epochs. (b) demonstrates strong robustness on the imbalanced dataset. These results indicate that our model is efficient and robust, making it well-suited for sleep staging tasks.