Fig. 6

A. EEG recordings (1) are used to compute latent dynamical factors by either multi-frequency decomposition and nonlinear feature calculation (A.2 and A.3) or reservoir computing (B.3). Output from either method is input to a (4) supervised tensor factorization algorithm to extract latent dynamical features or “hidden neurophysiology.” Other available data, such as putative causes, patient history, or lab tests, may be input as covariates to supervised tensor factorization. Resulting latent dynamical factors are input to learning models, as in Fig. 1.