Unsupervised Representation Learning Generates Differentiable Neurophysiological Profiles
Human brain activity contains stable, individual-specific features that persist over months to years, forming neurophysiological profiles. Most model-based profiling approaches use participant labels or supervised objectives, making it difficult to determine whether successful differentiation reflects stable biology or exploitable idiosyncrasies. We introduce a participant-agnostic autoencoder framework that derives profiles from brief resting-state magnetoencephalography (MEG) segments using reconstruction as sole training objective. Discriminative profiles emerged from the learned latent space without participant labels. Within-session, autoencoder profiles reached 93.3% accuracy at 120 s, exceeding functional-connectivity, spectral, and contrastive baselines with recordings as short as 14 s when participant-specific anatomy was withheld from source reconstruction. Differentiation generalized above chance across recording sessions (between-session accuracy 49.5% for the pretrained autoencoder). Profiles also predicted age more accurately than baselines (r2=0.318), and the decoder enabled perturbation-based sensitivity analyses in spectral and connectivity spaces. This establishes participant-agnostic representation learning as a scalable and interpretable profiling.