Bayesian Nonnegative CP Decomposition-Based Feature Extraction Algorithm for Drowsiness Detection.

Area of Science:

• Neuroscience
• Signal Processing
• Machine Learning

Background:

• Daytime naps involve complex physiological states like alertness and drowsiness.
• Understanding the relationship between drowsiness and physiological signals is crucial for sleep studies.
• Electroencephalogram (EEG) signals offer insights into brain activity during sleep and wakefulness.

Purpose of the Study:

• To propose a novel Bayesian nonnegative CP decomposition (BNCPD) model for extracting multiway features from group-level EEG signals.
• To automatically determine the underlying CP rank using a Bayesian nonparametric approach.
• To evaluate the effectiveness of BNCPD-based features for drowsiness detection during daytime naps.

Main Methods:

• Developed a Bayesian nonnegative CP decomposition (BNCPD) model incorporating prior distributions for factor matrices.
• Applied variational inference for efficient approximation of posterior distributions.
• Compared BNCPD-derived features against traditional methods for drowsiness detection using EEG data.

Main Results:

• Simulations demonstrated the BNCPD model's accuracy in recovering the true CP rank.
• Experimental results showed BNCPD features outperformed traditional methods in drowsiness detection accuracy.
• The model proved effective across various parameter settings.

Conclusions:

• The BNCPD model offers a robust method for automatic CP rank determination in multiway data.
• This approach provides valuable multiway physiological information for understanding individual states during naps.
• BNCPD is a promising tool for enhancing drowsiness detection and sleep research.