TCPL: task-conditioned prompt learning for few-shot cross-subject motor imagery EEG decoding

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Motor imagery (MI) electroencephalogram (EEG) decoding is crucial for brain-computer interfaces (BCIs).
• Significant inter-subject variability and limited data hinder few-shot cross-subject adaptation in existing MI-EEG decoding methods.
• Current approaches often require extensive fine-tuning or fail to capture individual neural patterns.

Purpose of the Study:

• To develop an effective few-shot cross-subject adaptation method for MI-EEG decoding.
• To address the limitations of existing approaches in handling inter-subject variability and limited training data.
• To enhance the development of personalized brain-computer interface systems.

Main Methods:

• Proposed Task-Conditioned Prompt Learning (TCPL) integrating a Task-Conditioned Prompt (TCP) module with a hybrid Temporal Convolutional Network (TCN) and Transformer backbone.
• Employed a meta-learning framework for rapid adaptation with minimal samples.
• TCP module encodes subject-specific variability as prompt tokens; TCN extracts local temporal patterns; Transformer captures global dependencies.

Main Results:

• TCPL demonstrated strong generalization and efficient adaptation across unseen subjects on three public datasets (GigaScience, Physionet, BCI Competition IV 2a).
• The model effectively handles inter-subject variability and limited training data.
• Achieved robust performance in few-shot EEG decoding scenarios.

Conclusions:

• TCPL is a feasible approach for practical few-shot EEG decoding.
• The proposed method shows significant potential for advancing personalized brain-computer interface systems.
• Highlights the effectiveness of integrating prompt learning with hybrid deep learning architectures for BCI applications.