Group-level brain decoding with deep learning.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Brain imaging data decoding is crucial for brain-computer interfaces and understanding neural representations.
• Subject-specific decoding models often lack generalizability due to high between-subject variability.
• Overcoming this variability is key for richer neuroscientific insights and superior group-level models.

Purpose of the Study:

• To develop a novel method for decoding brain imaging data that effectively handles between-subject variability.
• To adapt the WaveNet architecture with subject embedding for improved classification accuracy in group-level decoding models.
• To investigate the impact of subject embedding on the performance gap between subject-specific and group-level decoding.

Main Methods:

• A WaveNet-based deep learning architecture was adapted for classification tasks.
• Subject embedding, analogous to natural language processing word embeddings, was employed to model between-subject variability.
• The method was applied to magnetoencephalography (MEG) data from 15 subjects viewing 118 images.

Main Results:

• The combination of deep learning and subject embedding significantly reduced the performance gap between subject- and group-level decoding models.
• Group models demonstrated superior performance on low-accuracy subjects and could aid in initializing subject models.
• Permutation feature importance provided insights into the spatiotemporal and spectral information utilized by the group models.

Conclusions:

• Subject embedding is a crucial component for enhancing group-level decoding models in brain imaging.
• This approach offers a promising avenue for more generalizable and robust brain decoding, particularly with larger datasets.
• The study provides a framework for group-level interpretation of decoding models, offering physiological insights.