Uncovering the spatiotemporal structure of neural avalanches through optimal transport and dynamic time warping

Neural avalanches--or threshold-defined bursts of coordinated activity--are traditionally characterised by scale-free statistics. However the study of their detailed spatiotemporal structure and their recurrence within spontaneous activity has been hindered by the variability of avalanches in duration and spatial extent. To tackle this challenge, we propose the use of flexible alignment: we employ a distance metric combining unbalanced optimal transport with subsequence dynamic time warping, enabling comparison across events of different lengths and spatial configurations. Applied to 64-channel EEG from 63 participants in the PsiConnect psilocybin study, hierarchical clustering revealed 12 recurring propagation patterns. These cluster templates were then traced in the original continuous recordings, identifying sequences where the same pattern recurred consecutively and immediately at least twice. Sequences with alternating polarity were classified as oscillating; those with consistent polarity as stable. Oscillating sequences predominantly corresponded to clusters exhibiting visually confirmed spatial propagation, while stable sequences corresponded to spatially fixed patterns. Under psilocybin, oscillating sequences were reduced relative to stable sequences, shifting the polarity balance toward stable; this overall shift was confirmed by a subject-level permutation test, while the apparent task- and training-specific effects did not survive it. This effect was also mostly driven by a specific cluster, suggesting that there is concrete neural dynamics that are temporally affected by the consumption of psilocybin. The developed methodology has been implemented in a publicly available stppy Python package. Author SummaryWhen large groups of neurons activate together in brief bursts--called neural avalanches--they create patterns that sweep across the brains surface. These patterns are usually studied only as statistics (how big, how frequent), giving little attention to what they actually look like. We developed a method to compare the shapes of individual avalanches, group similar ones together, and then use the resulting templates as detectors to track when each pattern reappears in continuous brain recordings. Applied to high-density EEG data from 63 human subjects, we discovered 12 propagation patterns, and analysed whether they oscillate (i.e., repeat with alternating polarity) or remain stable, maintaining a fixed spatial arrangement. We found that during psilocybin intoxication, oscillating patterns were suppressed relative to stable patterns, making stable patterns relatively more prevalent in the detected sequences. This overall shift was statistically reliable across participants, although the apparent differences between individual tasks were not. Our approach bridges three previously separate research areas--avalanche dynamics, EEG microstates, and travelling waves--suggesting they describe the same underlying brain activity from different angles: some of our clusters resemble canonical microstate topographies, while others reveal the travelling wave dynamics that microstate analysis, by design, cannot sufficiently capture.