Deep Learning of Explainable EEG Patterns as Dynamic Spatiotemporal Clusters and Rules in a Brain-Inspired Spiking...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models often lack interpretability, especially when analyzing complex spatiotemporal brain data (STBD).
• Spiking Neural Networks (SNNs) offer a brain-inspired approach but require methods to enhance their explainability.
• Incremental and on-line learning from streaming data presents challenges in knowledge discovery.

Purpose of the Study:

• To develop a new SNN architecture for deep learning and knowledge discovery from streaming spatiotemporal brain data (STBD).
• To enhance the explainability of SNN models through dynamic neural clustering and the extraction of spatiotemporal rules.
• To improve feature selection and classification accuracy by identifying important STBD features and corresponding brain regions.

Main Methods:

• Proposed a brain-inspired Spiking Neural Network (SNN) architecture for incremental and on-line learning.
• Developed a dynamic neural clustering method to visualize and analyze SNN learning processes, capturing clusters as evolving polygons.
• Quantitatively analyzed cluster dynamics and spike-driven events to identify significant STBD features and activated brain regions.

Main Results:

• The SNN model extracted spatiotemporal rules, explaining prediction decisions and enhancing model interpretability.
• Dynamic neural clusters revealed important STBD features corresponding to specific brain regions.
• Application to Electroencephalogram (EEG) data from healthy and opiate-using subjects showed distinct cluster trends, enabling marker feature selection.
• Achieved 92% accuracy in EEG classification by using selected marker features, outperforming all-feature classification.

Conclusions:

• The proposed SNN method enhances the interpretability of learning behavior via dynamic neural clustering.
• The approach facilitates effective feature selection, leading to improved classification accuracy for STBD.
• Extracted spatiotemporal rules provide explainability for SNN model decisions.
• The study offers a deeper understanding of STBD dynamics, feature interactions, and brain region activity patterns.