Knowledge-driven feature component interpretable network for motor imagery classification.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Convolutional Neural Networks (CNNs) excel at motor imagery (MI) classification but lack interpretability and are computationally intensive.
• Existing deep learning models are purely data-driven, hindering knowledge discovery and specific network design.
• High computational costs limit the real-time application of CNNs in MI classification.

Purpose of the Study:

• To develop a novel knowledge-driven, interpretable network (KFCNet) for motor imagery classification.
• To address the limitations of existing purely data-driven deep learning models.
• To enable real-time MI classification with high performance and reduced computational cost.

Main Methods:

• KFCNet integrates spatial and temporal convolutions, initialized with prior frequency band knowledge of sensory-motor rhythms via band-pass filters.
• A symmetry loss function is introduced to ensure linear phase and unimodality of filters, alongside cross-entropy loss.
• Subject-specific time-frequency properties of event-related desynchronization/synchronization are used to construct and initialize the network with fewer parameters.

Main Results:

• Experiments on two public datasets demonstrated interpretable feature components in the trained KFCNet model.
• The physically meaningful features facilitated efficient network structure design.
• KFCNet achieved excellent classification performance on motor imagery tasks.

Conclusions:

• KFCNet offers comparable performance to state-of-the-art methods while utilizing significantly fewer parameters.
• The interpretability of KFCNet facilitates knowledge discovery and network design.
• The reduced computational load makes KFCNet suitable for real-time motor imagery applications.