Convolutional neural networks decode finger movements in motor sequence learning from MEG data.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Non-invasive Brain-Computer Interfaces (BCIs) accurately classify hand lateralization.
• Distinguishing individual finger movements is challenging due to overlapping motor cortex representations.

Purpose of the Study:

• Validate a compact convolutional neural network (CNN) for decoding finger movements from magnetoencephalography (MEG).
• Assess the performance and interpretability of the Linear Finite Impulse Response Convolutional Neural Network (LF-CNN) against other deep learning models.

Main Methods:

• Recorded MEG data from participants performing a serial reaction time task (SRTT) involving index and middle finger presses.
• Developed and compared LF-CNN with EEGNet, FBCSP-ShallowNet, and VGG19 for classifying hand and finger movements.

Main Results:

• All models achieved >95% accuracy for hand laterality decoding.
• Individual finger movement decoding accuracy ranged from 80-85%.
• LF-CNN demonstrated superior computational efficiency and interpretability in spatial and spectral domains.

Conclusions:

• A tailored CNN (LF-CNN) enables feasible and accurate decoding of individual finger movements from non-invasive MEG.
• LF-CNN offers comparable performance to complex architectures with faster, interpretable results.
• This approach has significant potential for investigating neural mechanisms in cognitive neuroscience.