Extended Data Fig. 1. Computational experiment setup for all candidate SOZ features and statistical analysis.

Computational experiment setup for all candidate SOZ features and statistical analysis - (a) Any candidate feature that can produce a spatiotemporal heatmap was computed from EEG data and then partitioned by the clinically annotated SOZ set and the complement, SOZ^C (i.e. non-SOZ electrodes) to compute a confidence statistic measuring the feature’s belief of the clinician’s hypothesis. Here F_SOZ and F_SOZ^C were the feature values within their respective sets. f_θ is the function depending on the Random Forest model parameters, θ that maps the statistics of the F_SOZ and F_SOZ^C to a confidence statistic. An ideal feature would have high and low confidence for success and failed outcomes respectively. Each point on the final CS distribution comparisons represent one patient. (b) A more detailed schematic of how our proposed fragility and baseline features were computed from EEG data for a single snapshot of EEG data. See fragility methods section for description of x, A and Δ.