Combining Statistical Analysis and Machine Learning for EEG Scalp Topograms Classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain-computer interfaces (BCIs) require extensive user calibration, limiting their daily application.
• Developing pre-trained decoders can significantly reduce user burden and calibration time.

Purpose of the Study:

• To develop a pre-trained artificial neural network classifier for brain-computer interfaces (BCIs) that demonstrates high accuracy on naive users.
• To address the cross-subject variability challenge in BCI decoding through transfer learning.

Main Methods:

• Trained an artificial neural network on ambiguous stimuli classification tasks using time-frequency features from electroencephalographic (EEG) spectral power.
• Extracted shared neurophysiological features from a representative group of subjects to build a pre-trained classifier.
• Statistically contrasted EEG spectral power between low and high ambiguity stimuli classes.

Main Results:

• The pre-trained classifier achieved 74% accuracy on data from newly recruited, naive subjects.
• The method effectively utilizes fundamental neurophysiological processes shared across individuals.
• Identified interpretable feature subspaces for transfer learning in cross-subject BCI tasks.

Conclusions:

• Pre-trained BCI decoders can bypass initial training, improving usability for individuals with paralysis or limb amputation.
• This approach facilitates transfer learning for BCI, mitigating negative transfer in cross-subject applications.
• Contributes to making BCIs more accessible and efficient for diverse user populations.