Unsupervised Sleep and Wake State Identification in Long-Term Electrocorticography Recordings

Area of Science:

• Neuroscience
• Computational Neuroscience
• Sleep Medicine

Background:

• Sleep research traditionally relies on manual annotation of sleep stages.
• Automating this process is crucial for understanding atypical sleep patterns and neurological disorders.
• Existing methods may struggle with complex or non-standard sleep data.

Purpose of the Study:

• To develop a fully unsupervised approach for labeling sleep and wake states.
• To evaluate the performance of hidden semi-Markov models (HSMM) against other methods using human electro-corticography (ECoG) data.
• To enable more efficient and accurate sleep research, particularly in patient populations.

Main Methods:

• Utilized continuous electro-corticography (ECoG) data from a single electrode in epilepsy patients.
• Implemented and compared hidden semi-Markov models (HSMM), K-means clustering, and hidden Markov models (HMM).
• Focused on unsupervised classification of sleep and wake states without manual pre-labeling or excessive state transitions.

Main Results:

• Hidden semi-Markov models (HSMM) demonstrated superior performance with a mean accuracy of 85.2%.
• HSMM outperformed K-means clustering (72.2%) and hidden Markov models (81.5%).
• The HSMM approach effectively classified sleep/wake states with minimal spurious transitions.

Conclusions:

• HSMMs provide a robust and accurate method for unsupervised sleep/wake state annotation in ECoG data.
• This unsupervised approach facilitates sleep research and understanding of neurological conditions.
• The developed model serves as a foundation for clustering sleep stages and other behavioral states.