Unsupervised Representation Learning Generates Differentiable Neurophysiological Profiles.

Recent neuroimaging research has shown that human brain activity expresses stable, individual-specific features that persist over months to years, defining neurophysiological profiles. Current model-based profiling relies on labeled data and supervised learning, leaving open whether they exploit idiosyncratic artifacts or genuine biology. We introduce a participant-agnostic autoencoder framework to derive differentiable profiles from brief segments of resting-state magnetoencephalography (MEG). Despite an unsupervised objective, discriminative profiles emerged naturally from the learned latent space, outperforming model-free and model-based baselines in participant differentiation. Reliable differentiation was achieved using recordings as short as 14 s, generalized across recording sessions, and remained robust without anatomical information. Beyond differentiation, learned profiles predicted age more accurately than baselines, and the decoder enabled perturbation-based sensitivity analyses directly in spectral and connectivity spaces. These results establish participant-agnostic modeling as a principled and interpretable framework for neurophysiological profiling that generalizes across sessions while preserving sensitivity to biologically relevant individual differences.