Brain decoding of the Human Connectome Project tasks in a dense individual fMRI dataset

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning

Background:

• Brain decoding aims to interpret cognitive states from brain activity patterns.
• Significant individual differences in brain organization hinder accurate group-level decoding.
• The potential for accurate individual-level brain decoding remains largely unexplored.

Purpose of the Study:

• To investigate the feasibility of training accurate brain decoding models solely at the individual level.
• To evaluate various machine learning methods for individual brain decoding using dense fMRI data.
• To establish a benchmark for future research on cross-subject and cross-condition model generalization.

Main Methods:

• Trained nine decoding models, including SVM, MLP, and GCN, on individual fMRI data from six participants.
• Utilized dense, subject-specific datasets (approx. 7 hours per participant) from the Human Connectome Project (HCP) task battery.
• Classified single fMRI volumes into 21 experimental conditions simultaneously.

Main Results:

• Graph Convolutional Networks (GCN) and Multi-Layer Perceptrons (MLP) achieved the highest accuracies (57-67%), approaching group-level performance.
• Support Vector Machines (SVM) also demonstrated strong performance (54-62%) in individual brain decoding.
• Feature importance maps highlighted domain-specific cognitive regions, particularly in the motor cortex, and confirmed individual-specific feature learning.

Conclusions:

• Densely sampled, individual neuroimaging datasets enable the training of accurate brain decoding models.
• Individual-level decoding models show superior performance compared to inter-subject classification, underscoring the importance of subject-specific features.
• This study provides a valuable benchmark for developing advanced decoding techniques that generalize across subjects and acquisition parameters.