MCL

MCL: Attention-Driven Meta-Learning for Robust Cross-Subject EEG-fNIRS Classification

This repository contains the implementation of Meta-CNN-LSTM (MCL), an attention-driven meta-learning model designed for robust cross-subject classification of EEG-fNIRS data. The MCL model integrates:

• Meta-learning optimization for adaptive subject generalization.
• CNN-LSTM hybrid architecture to capture spatial-temporal dependencies.
• Attention mechanisms to enhance feature importance.
• Event-Related Desynchronization (ERD/ERS)
• Feature optimization across time (TD), frequency (FD), and time-frequency (TFD) domains.

MCL achieves 96.6% cross-subject accuracy on cognitive BCI tasks, outperforming state-of-the-art methods while maintaining efficiency for real-time brain-computer interface (BCI) applications.

Features

• Meta-Learning Framework: Adaptively optimizes feature representations for subject-independent classification.
• EEG-fNIRS Data Fusion: Integrates complementary neural signals to enhance classification robustness.
• Attention Mechanism: Selectively enhances critical EEG-fNIRS features, reducing irrelevant noise.
• Optimized Feature Selection: Identifies key EEG-fNIRS features (e.g., FDS, TDT) for improved model generalization.

File Structure

• DataAugmentation.py --> Data augmentation pipeline for EEG-fNIRS features
• ModelInitialization.py --> Initializes task distributions and optimizers
• TrainingLoop.py --> Implements inner and outer loop training for meta-learning
• ModelArchitecture.py --> Defines the CNN-LSTM architecture with attention mechanism

If you find this work useful, please cite our article: