Sleep Stage Classification Using Time-Frequency Spectra From Consecutive Multi-Time Points.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Data Science

Background:

• Sleep stage classification is a complex challenge in sleep research.
• Previous studies often utilize limited datasets.
• Accurate sleep staging is crucial for diagnosing sleep disorders.

Purpose of the Study:

• To evaluate the effectiveness of Long Short-Term Memory (LSTM) networks for sleep stage classification.
• To compare LSTM performance against Convolutional Neural Networks (CNNs).
• To investigate the impact of temporal information on classification accuracy.

Main Methods:

• Utilized the Sleep Heart Health Study dataset.
• Employed an LSTM network with time-frequency spectra from consecutive 30s intervals.
• Compared LSTM with four classical CNNs using single 30s time points.

Main Results:

• LSTM networks demonstrated superior performance compared to CNNs.
• Classification accuracy increased with the inclusion of additional temporal information.
• Peak performance achieved with three consecutive 30s time points (87.4% accuracy, 0.8216 Cohen's Kappa).

Conclusions:

• Temporal information is critical for accurate sleep stage classification.
• LSTM networks are well-suited for sleep stage classification due to their ability to process temporal data.
• Findings suggest LSTM networks may be more appropriate than CNNs for this task.