Multidomain Convolution Neural Network Models for Improved Event-Related Potential Classification.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Event-related potentials (ERPs) are crucial neural signals for understanding brain activity.
• Accurate classification of ERPs is essential for applications like brain-computer interfaces (BCIs) and diagnosing brain disorders.
• Existing methods often struggle with the complexity and variability of ERP data across subjects and trials.

Purpose of the Study:

• To introduce two novel convolution neural network (CNN) models for accurate ERP classification.
• To fuse frequency, time, and spatial domain information from continuous wavelet transform (CWT) of ERPs.
• To evaluate the models' performance in both individual (BCI) and group-based (disorder classification) scenarios.

Main Methods:

• Developed two multidomain CNN models utilizing Z-scalograms and V-scalograms derived from CWT.
• Fused multichannel ERP data into frequency-time-spatial cuboids (first model) and matrices (second model).
• Conducted experiments for customized (single-subject) and group-based (cross-subject) ERP classification.

Main Results:

• Both multidomain models achieved high classification accuracies for single trials and averaged ERPs.
• Effective classification was demonstrated even with a small subset of top-ranked channels.
• The proposed multidomain fusion models consistently outperformed unichannel classifiers.

Conclusions:

• The novel multidomain CNN models effectively leverage fused frequency, time, and spatial information for ERP classification.
• These models show significant promise for both BCI applications and brain disorder classification.
• The fusion approach offers a superior alternative to traditional unichannel methods for ERP analysis.