Unsupervised Representation Learning Generates Differentiable Neurophysiological Profiles.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Human brain activity exhibits stable, individual-specific features over time, termed neurophysiological profiles.
• Current profiling methods often require labeled data and supervised learning, raising questions about their reliance on genuine biological signals versus artifacts.

Purpose of the Study:

• To develop a participant-agnostic autoencoder framework for deriving differentiable neurophysiological profiles from resting-state magnetoencephalography (MEG) data.
• To assess the performance of unsupervised learning in identifying genuine biological individual differences in brain activity.

Main Methods:

• Implemented a participant-agnostic autoencoder framework to process brief segments of resting-state MEG data.
• Utilized an unsupervised learning objective to derive profiles from the latent space.
• Compared the developed framework against model-free and model-based baselines for participant differentiation and age prediction.

Main Results:

• Unsupervised learning naturally yielded discriminative neurophysiological profiles that outperformed baseline methods in participant differentiation.
• Reliable differentiation was achieved with short recordings (14s), generalized across sessions, and was robust without anatomical information.
• Learned profiles predicted age more accurately than baselines, and the framework allowed for sensitivity analyses in spectral and connectivity spaces.

Conclusions:

• Participant-agnostic modeling provides a principled, interpretable framework for neurophysiological profiling.
• This approach generalizes across recording sessions and effectively captures biologically relevant individual differences.
• The method offers a robust alternative to supervised learning for identifying unique brain activity signatures.