Structured recurrence of latent neural processes predicts memory recognition performance

Abstract

Understanding cognition requires explaining not only which neural processes are engaged, but how their relationships evolve through time. We tested whether the recurrent organization of low-dimensional whole-brain dynamics relates to auditory recognition behaviour beyond activation magnitude. After a brief learning phase, 65 participants underwent magnetoencephalography (MEG) while recognizing three-, five-, and seven-sound auditory sequences, presented unchanged or with a single early or late tone alteration. Broadband Brain Network Estimation via Source Separation (BROAD-NESS) identified broadband latent networks from source-reconstructed activity and represented their joint activation as trajectories through a multidimensional state space; recurrence quantification analysis characterized their temporal organization. Memorized melodies elicited more recurrent states and longer, more deterministic and persistent recurrence structures, together with lower divergence, than changed melodies. Across participants, structured recurrence was consistently associated with higher accuracy and faster responses, whereas recurrence frequency showed no consistent behavioural association. After adjustment for condition, melody length, their interaction, and matched root-mean-square activation amplitude, recurrence measures increased marginal R2 by 0.273. Nested participant-wise cross-validation reduced held-out prediction error by 11.81% (95% CI: 4.01%-17.65%). These findings support a view of cognition in which performance depends critically on how low-dimensional neural processes jointly organize and evolve over time, beyond how strongly they are activated.