Exploiting the Cone of Influence for Improving the Performance of Wavelet Transform-Based Models for ERP/EEG...

Area of Science:

• Signal Processing
• Machine Learning
• Neuroscience

Background:

• Wavelet transform features are crucial for classifying electroencephalography (EEG) and event-related potential (ERP) signals.
• Edge effects in continuous wavelet transform (CWT) introduce unreliable features in scalograms, impacting classifier performance.

Purpose of the Study:

• To enhance wavelet-based machine learning classifiers for EEG/ERP signals by effectively utilizing the cone of influence (COI).
• To investigate the impact of processing CWT scalograms by zeroing out or cropping features outside the COI on classification accuracy.

Main Methods:

• Extracted features from three types of scalograms: standard (S), zeroed-out (Z), and valid/cropped (V).
• Developed subsampling strategies for small-sample ERP ensembles and channel selection for multi-channel data.
• Implemented and compared various machine learning models including SVM, Random Forests, KNN, MLP, and CNN on ERP data from subjects distinguishing word pairs.

Main Results:

• Classifiers utilizing Z-scalogram features outperformed those using S-scalogram features.
• Classifiers using V-scalogram features demonstrated superior performance compared to Z-scalogram features.
• A dramatic relative improvement in performance was observed for V-scalogram classifiers over standard S-scalogram classifiers.

Conclusions:

• Cropping unreliable features outside the COI (V-scalogram) significantly improves EEG/ERP classification accuracy.
• The proposed method enables customized classifier design for individual subjects, advancing patient-specific diagnoses.
• This approach offers a substantial performance boost for wavelet-based signal classification in neuroscience and clinical applications.