Application of Transformers for Sleep Stage Classification

Abstract

[Abstract]: Transformer architectures have revolutionized the field of artificial intelligence in recent years. This work investigates the application of Transformers to sleep stage classification, a complex medical problem that remains an active area of research. Despite promising results in recent studies, the clinical application of these methods remains limited. In this paper, we propose a novel model that replaces the LSTM component of a state-of-the-art CNN+LSTM architecture with a Transformer encoder. We compare the performance of our proposed approach against our baseline and several successful state-of-the-art models. The results demonstrate several advantages for our proposed method, including faster convergence during training, reduced model complexity, and enhanced performance. Leveraging the Transformer architecture, we also propose and investigate an interpretability method based on attention score analysis. Finally, we evaluate the inter-database generalization performance of our proposed model using an external database.