MI-LTN: A Neurosymbolic Framework for Enhanced EEG Feature Extraction and Model Interpretability in MI-BCI.

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Brain-Computer Interface (BCI) technology enables human-computer interaction.
• Motor Imagery Electroencephalogram (MI-EEG) decoding is a key area in BCI research.
• Deep learning for EEG signal decoding faces challenges in feature representation, extraction, and interpretability.

Purpose of the Study:

• To propose a novel neurosymbolic framework, MI-LTN, for improved MI-EEG decoding.
• To address challenges in feature representation, extraction, and interpretability in BCI.
• To incorporate logical constraints and channel importance evaluation into deep learning models.

Main Methods:

• Developed a neurosymbolic framework named MI-LTN (Motor Imagery Logic Tensor Network).
• Integrated logical constraints using Logic Tensor Network (LTN) into the training model.
• Utilized Shapley values for evaluating and adjusting channel importance.

Main Results:

• Achieved 86.00% classification accuracy on the BCI IV 2a dataset.
• Achieved 88.84% classification accuracy on the BCI IV 2b dataset.
• Demonstrated the effectiveness of the neurosymbolic approach in MI-EEG decoding.

Conclusions:

• The proposed MI-LTN framework shows significant potential for advancing MI-EEG decoding in BCI.
• Logic Tensor Networks (LTN) offer a promising approach for enhancing feature representation and interpretability.
• The method effectively addresses key challenges in current deep learning-based EEG signal processing.