MEG Neural Decoding Pipeline: The Issues Residing Within The Data and Methods to Improve Your Decoding Accuracy.

Area of Science:

• Neuroscience
• Data Science
• Machine Learning

Background:

• Analyzing magnetoencephalography (MEG) data for neural decoding is challenging with small datasets.
• Classic stationary data analysis methods can introduce distortions.
• Existing data augmentation techniques often rely on synthetic data, limiting neuroscientific applicability.

Purpose of the Study:

• To develop a robust analysis pipeline for neural decoding of small MEG datasets.
• To introduce a novel data augmentation technique applicable to real neuroscientific data.
• To enhance the accuracy of machine learning algorithms in neural decoding tasks.

Main Methods:

• A hybrid analysis pipeline combining stationary and non-stationary methods for MEG data.
• Application of Fourier-based methods, empirical mode decomposition, and principal component analysis for data cleaning.
• A novel data augmentation technique using combinations' averaging on real data.
• Automated epoch rejection and comparison of four machine learning designs.

Main Results:

• The proposed pipeline effectively cleans MEG data and compensates for distortions.
• The combinations' averaging data augmentation technique is reliable and does not create unnatural patterns.
• Significant increases in neural decoding accuracy were achieved through channel selection, feature dimension reduction, and epoch averaging.
• Machine learning models demonstrated improved performance on small-sized datasets.

Conclusions:

• The developed analysis pipeline enables effective neural decoding of small MEG datasets.
• The novel data augmentation method broadens the scope of machine learning applications in neuroscience.
• Optimized feature selection and data processing significantly enhance decoding accuracy, making advanced analyses accessible for limited data.