Sleep Staging Using Plausibility Score: A Novel Feature Selection Method Based on Metric Learning.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Biomedical Signal Processing

Background:

• Feature selection is crucial for reducing complexity and enhancing classification performance in machine learning.
• Existing feature selection methods often rely on Euclidean distance, limiting their effectiveness with complex data like pairwise constraints.
• Metric learning offers a promising approach to jointly evaluate feature subsets, but effective criteria for pairwise constraints are needed.

Purpose of the Study:

• To propose a novel filter method for feature selection using pairwise constraints and metric learning.
• To design criteria based on Kullback-Leibler divergence to evaluate feature subset discrimination.
• To improve classification performance in electroencephalogram (EEG) based sleep staging.

Main Methods:

• A filter method for feature selection utilizing pairwise constraints and metric learning.
• Two criteria based on Kullback-Leibler divergence to measure differences in must-link and cannot-link constraints.
• Sequential search algorithm guided by simplified criteria derived from Kullback-Leibler divergence.
• Application to sleep staging using electroencephalogram (EEG) data from the Sleep-EDF Database Expanded.

Main Results:

• The proposed method effectively selects features by jointly evaluating subsets based on metric learning.
• Experimental results on EEG sleep staging demonstrate superior performance compared to nine representative feature selection methods.
• Achieved high average accuracies of 97.66% for healthy volunteers and 93.57% for individuals with mild sleep difficulties.

Conclusions:

• The proposed feature selection method, incorporating metric learning and Kullback-Leibler divergence, is effective for EEG-based sleep staging.
• This approach offers significant improvements in classification accuracy and computational efficiency.
• The method shows promise for analyzing complex biomedical signals with pairwise constraints.