Unsupervised Feature Representation of Sleep EEG Data with Transient Deep Boltzmann Machine

Area of Science:

• Computational neuroscience
• Machine learning for healthcare
• Sleep science

Background:

• Supervised sleep staging demands extensive labeled datasets, posing a significant challenge.
• Unsupervised methods offer a potential solution to overcome data limitations in sleep analysis.

Purpose of the Study:

• To develop and evaluate an unsupervised dimensionality reduction technique for binary sleep stage classification.
• To assess the performance of a transiently trained Deep Boltzmann Machine (DBM) against other methods.

Main Methods:

• Extracted joint time-frequency domain features from polysomnographic recordings.
• Smoothed features using a 2-minute rolling window for temporal context.
• Applied unsupervised training of a Deep Boltzmann Machine (DBM) to a transient state.

Main Results:

• The DBM_transient method effectively separated sleep stages in a 2D feature space, indicated by a high Fisher's discriminant value.
• Achieved a 96.1% F1 score, outperforming converged DBM (95.2%) and other dimensionality reduction techniques.
• Demonstrated superior classification performance compared to Principal Component Analysis, Isometric Feature Mapping, t-SNE, and UMAP.

Conclusions:

• The transient Deep Boltzmann Machine (DBM_transient) offers a highly effective unsupervised approach for sleep stage classification.
• The method's interpretability in 2D space is a significant advantage.
• Future multi-class implementations could enhance clinical applicability of this sleep analysis technique.