Machine Learning on Magnetoencephalography Data Yields Generalizable Low-Dimensional Neural Fingerprints That...

Area of Science:

• Neuroscience
• Computational Neuroscience
• Brain Imaging

Background:

• Individual brains exhibit unique structural and functional characteristics.
• Neural fingerprints, reflecting functional connectivity, capture individual differences in behavior and cognition.
• High-dimensional functional connectomes pose computational challenges for machine learning in neuroimaging.

Purpose of the Study:

• To develop a low-dimensional alternative for capturing individual features in brain functional connectivity.
• To evaluate the performance of latent-noise Bayesian Reduced Rank Regression (lnBRRR) for creating neural fingerprints.
• To assess if individual features captured by lnBRRR are altered by different cognitive processes.

Main Methods:

• Employed latent-noise Bayesian Reduced Rank Regression (lnBRRR) on MEG-derived functional connectivity and power spectral density data.
• Assessed lnBRRR performance with small training set sizes (N=20-44) and compared it against principal component analysis and linear discriminant analysis.
• Evaluated model performance using task data and compared solutions across conditions to identify task-related alterations in individual features.

Main Results:

• lnBRRR captured generalizable individual patterns with as few as 20 participants, with optimal accuracy achieved at 30-35 participants.
• The model demonstrated comparable performance to alternative dimensionality reduction techniques.
• Latent fingerprints derived from task data performed similarly to resting-state fingerprints, indicating generalizability across conditions.
• Individual differences in power spectral density were found to be largely intrinsic and unaffected by cognitive task variations.

Conclusions:

• lnBRRR is a promising tool for analyzing neuroimaging data, offering a computationally efficient way to derive neural fingerprints.
• Individual differences in functional connectivity and power spectral density are robust and largely independent of specific cognitive tasks.
• The findings suggest that low-dimensional representations can effectively preserve unique individual brain characteristics for various applications.