We thank James Antony, Joseph McGuire, Chang-Hao Kao for sharing the data, Joshua Faskowitz & the brain networks & behavior lab ( www.brainnetworkslab.com ) for sharing the edge time series code. Thanks to @monicarosenb.bsky.social for the help and support on this project. 9/9

Nor did models built from related behavioral measures (e.g., participants’ prediction, reward). 8/9

Moreover, models built from BOLD activation alone failed to generalize across datasets to predict surprise. 7/9

The data-driven surprise EFPM outperformed models built from interactions between and/or within predefined functional brain networks. 6/9

The same model generalized to predict surprise when people watched NCAA basketball games (www.sciencedirect.com/science/arti...), even when controlling for other features in the games (e.g., video motion). 5/9

The model predicted surprise in the adaptive learning task (www.sciencedirect.com/science/arti...) in held-out individuals from their functional network dynamics. 4/9

Using insights from edge-centric neuroscience (www.nature.com/articles/s41...), we built an edge-fluctuation-based predictive model (EFPM) to identify functional interactions predicting moment-to-moment changes in surprise in an adaptive learning task. 3/9

This is difficult to assess with behavioral measures alone because in some paradigms surprise is measured explicitly whereas in others it is hidden. Characterizing brain dynamics allows us to discover commonalities between surprise in different contexts. 2/9