Enhancing sleep stage classification through simultaneous time-frequency tokenization

doi:10.1016/j.bspc.2025.107553

Enhancing sleep stage classification through simultaneous time-frequency tokenization

2025-02-28

0

PRE

AI

DOI：10.1016/j.bspc.2025.107553

Original

Original request for help

Save

Abstract

En 中文

Accurate sleep stage classification is crucial for assessing sleep quality and diagnosing sleep disorders. The current sleep stage classification methods based on time-frequency images generally adopt Transformer architecture design, taking the time dimension or frequency dimension as the token, which will make it impossible to extract the time domain and frequency domain information at the same time. In order to address this issue, this paper proposes a novel Multi-branch Sleep Stage Classification Network (MBSleepNet), which consists of three branches: Time Branch, Frequency Branch, and Global Branch. The Time Branch and Frequency Branch are designed based on the Transformer, treating the time and frequency dimensions as tokens, respectively, to simultaneously extract time domain and frequency domain information. The Global Branch uses CNN to extract cross-modal global features from the time-frequency image. In addition, to fully utilize the information extracted by these three branches, we also designed a Multi-branch Information Fusion Module (MBIFM) to fully integrate time domain, frequency domain, and cross-modal global information to enhance sleep stage classification. Our method achieves a state-of-the-art result on the SleepEDF-20, SleepEDF78 and SHHS datasets. Specifically, MBSleepNet achieves 88.2% accuracy on the SleepEDF-20, 86.5% on the SleepEDF-78, and 89.2% on the SHHS.

Keywords:

Automatic sleep stage classification

Time-frequency image

Feature fusion

EEG

Enhancing sleep stage classification through simultaneous time-frequency tokenization

Abstract

Journal

Organization