P²CSL: cross-subject EEG classification by subspace class prototype-based progressive confident target sample...

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Domain adaptation (DA) is crucial for cross-subject electroencephalogram (EEG) decoding, addressing data distribution discrepancies.
• Existing methods face challenges with unreliable pseudo-labeling in early stages and balancing sample contributions later on.

Purpose of the Study:

• To propose a novel method, prototype-based progressive confident target sample labeling (P²CSL), for improved EEG decoding.
• To address the issues of error propagation from unreliable early pseudo-labels and the need for balanced sample contributions in DA.

Main Methods:

• P²CSL utilizes subspace class prototypes to aid in labeling target samples within a unified framework.
• It integrates domain-invariant EEG feature learning with self-supervised target sample labeling.
• Confident target samples are progressively incorporated into the DA model fitting process.

Main Results:

• P²CSL demonstrated competitive performance in cross-subject EEG classification tasks, including emotion recognition and inner speech decoding.
• The method outperformed state-of-the-art (SOTA) approaches in experiments.
• Fine-grained analyses confirmed the effectiveness of the sample confidence allocation strategy and dynamic model optimization.

Conclusions:

• The study highlights the effectiveness of considering both target sample reliability and their contribution to model training in DA.
• P²CSL offers a robust solution for enhancing cross-subject EEG decoding accuracy.
• The findings provide insights into stabilizing training and optimizing DA models through progressive sample incorporation.