Scholarlink

Sleep staging is vital for the examination of sleep quality and sleep disorders. The state-of-the-art process of manual sleep staging by experts is time-consuming. Therefore, reliable and accurate automatic sleep staging is of great value for clinical sleep medicine. In this work, the challenge of developing a reliable automated sleep staging algorithm using single-channel Electroencephalography (EEG) data is tackled. This is achieved through a hybrid model, which combines the SleePyCo sleep stage classifier, a transformer-encoder architecture, with a Hidden Markov Model (HMM). Utilizing labeled EEG data from 79 healthy subjects from the Sleep-EDF Expanded dataset, the parameters of the HMM are trained through backpropagation, using a Maximum-Mutual-Information-based loss function, and subsequently evaluated. The distinctive feature of this approach is the focus on the sequential properties of sleep stages. The results show that the hybrid model is able to improve overall performance compared to the state-of-the-art SleePyCo classifier. Moreover, it enhances interpretability, as the model’s transition probabilities are directly interpretable for expert users. This success demonstrates the potential of hybrid HMM-augmented classifiers for time-series analysis and pattern recognition. It also highlights challenges in this approach, laying groundwork for future developments.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Neuro-Explicit DNN-HMM Approach for Sleep Stage Classification: Enhancing Temporal Modeling and Interpretability

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Jakob Scheytt,
  • Eike Osmers,
  • Ludmila Litvin,
  • Dorothea Kolossa

摘要

Sleep staging is vital for the examination of sleep quality and sleep disorders. The state-of-the-art process of manual sleep staging by experts is time-consuming. Therefore, reliable and accurate automatic sleep staging is of great value for clinical sleep medicine. In this work, the challenge of developing a reliable automated sleep staging algorithm using single-channel Electroencephalography (EEG) data is tackled. This is achieved through a hybrid model, which combines the SleePyCo sleep stage classifier, a transformer-encoder architecture, with a Hidden Markov Model (HMM). Utilizing labeled EEG data from 79 healthy subjects from the Sleep-EDF Expanded dataset, the parameters of the HMM are trained through backpropagation, using a Maximum-Mutual-Information-based loss function, and subsequently evaluated. The distinctive feature of this approach is the focus on the sequential properties of sleep stages. The results show that the hybrid model is able to improve overall performance compared to the state-of-the-art SleePyCo classifier. Moreover, it enhances interpretability, as the model’s transition probabilities are directly interpretable for expert users. This success demonstrates the potential of hybrid HMM-augmented classifiers for time-series analysis and pattern recognition. It also highlights challenges in this approach, laying groundwork for future developments.