A knowledge discovery methodology from EEG data for cyclic alternating pattern detection.

Area of Science:

• Neurology
• Sleep Medicine
• Data Science

Background:

• Cyclic alternating patterns (CAP) detection and quantification hold potential as disease biomarkers.
• Existing methods for discriminating CAP components lack sensitivity and often use inappropriate accuracy measures for imbalanced datasets.

Purpose of the Study:

• To develop automatic CAP scoring approaches using knowledge discovery in data (KDD).
• To enable both binary classification (A-phases vs. B-phases) and multi-class classification of CAP components.
• To address the limitations of existing methods by employing a more robust performance metric.

Main Methods:

• A KDD methodology was applied, involving feature extraction (55 features), feature ranking/transformation (mRMR), and classification (SVM, k-NN, discriminant analysis).
• Weighted accuracy (WAC) was used to account for class imbalance, providing a more trustworthy performance assessment.
• The study utilized data from 30 subjects from the CAP Sleep Database.

Main Results:

• For multi-class CAP component discrimination, the best performance (WAC 51%) was achieved using mRMR feature ranking and SVM classification.
• For binary discrimination between A-phases and B-phases, k-NN with mRMR ranking yielded the highest WAC of 80%.
• The developed KDD approach demonstrated improved subtype sensitivities compared to previously published methods.

Conclusions:

• This KDD methodology represents a novel application to CAP scoring, particularly in fully discriminating the three A-phase subtypes.
• The use of WAC offers a more reliable performance evaluation for imbalanced CAP datasets.
• The study achieved superior subtype sensitivities, advancing the field of automated sleep analysis.