Fig. 8: Goodness-of-fit of pairwise MEM.

We measure the divergence between the empirically observed states’ empirical and estimated probabilities using a modified version of Kullback–Leibler divergence (see Materials and Methods section for details). a The relationship between the average (across all frequency bands, and recording montages) divergence of the empirical and estimated probabilities and the number of selected electrodes for each patient at ‘0’ (marked by ‘.’) and ‘1’ (marked by ‘x’) state binarization thresholds. The close fits of the linear regression models for both ‘0’ (slope = 0.55, p = 0.013, R² = 0.9) and ‘1’ (slope = 0.34, p = 0.003, R² = 0.96) thresholds suggest a negative relationship between the iEEG dimensions and the goodness-of-fit of pairwise MEM. b Bar plots show the average and standard deviation of divergence values for different recording montages and binarization thresholds (n = 5 patients × 5 frequency bands, except for the local montage states binarized at ‘0’ (n = 10) where the silent state (i.e., ‘all-off’) used for approximating the estimated probability function was observed only in 2 patients’ states). Statistical comparisons reveal that the divergence values are significantly smaller at higher binarization thresholds in the referential and the global montages (two-sided Wilcoxon rank-sum test, p < 0.05, marked by ‘*’). The divergence values are also lower in the referential montage compared to the other two montages only at the higher binarization threshold (two-sided Wilcoxon rank-sum test, p < 0.05, marked by ‘*’). c, d The normalized (i.e., z-scored across frequency bands) divergence values at different frequency bands for binarization thresholds of ‘0’ and ‘1’, respectively. Each dot represents a single patient, and their colors indicated the recording montage (color-coded similar to panel b). The open circles and error bars show the distributions’ mean and the standard deviation (n = 5 patients × 3 recording montages, with the above mentioned exception of local montage states binarized at ‘0’). Statistical comparisons reveal that the divergence values are significantly lower and higher at high γ compared to lower frequency bands using binarization thresholds of ‘0’ and ‘1’, respectively (two-sided Wilcoxon rank-sum test, p < 0.05, Bonferroni corrected for multiple comparisons, marked by ‘*’).