Subject-independent meta-learning framework towards optimal training of EEG-based classifiers

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Deep learning shows promise for Electroencephalography (EEG) signal classification.
• High inter-subject variability in EEG data hinders robust deep learning model training.
• Existing subject-adaptive methods require extensive target subject labels.

Purpose of the Study:

• To adapt model-agnostic meta-learning for subject-based EEG signal classification.
• To develop a few/zero-shot, subject-independent meta-learning framework.
• To achieve state-of-the-art performance in EEG classification tasks with reduced calibration.

Main Methods:

• Proposed a novel few/zero-shot subject-independent meta-learning framework.
• Applied the framework to multi-class inner speech and binary-class motor imagery classification.
• Evaluated model generalization and performance across subjects.

Main Results:

• Achieved significant improvements over current state-of-the-art methods.
• Binary class motor imagery classification reached 88.70% accuracy.
• Multi-class inner speech classification achieved an average accuracy of 31.15%.

Conclusions:

• The proposed framework effectively trains zero-calibration and few-shot models for subject-independent EEG classification.
• Demonstrated robust and generalized performance across subjects, performing well on datasets of varying sizes.
• Meta-learning methodology can be successfully modified for subject-based EEG signal classification.