Figure 5: Area under the curve and average precision performance.

Specific results for neural fragility are marked in red for each of the panels (a-d). (a) Discrimination plot (measured with AUC) shows the relative performance of benchmark feature representations compared to that achieved with neural fragility. Neural fragility has a median AUC of 0.89 (boxplot summary = 0.77, 0.97, 0.83, 0.93; min, max, first quartile, third quartile). (b) A similar average-PR curve shows the relative positive predictive value of all features compared with fragility. Average precision is the analagous area under the curve for the PR curve. Neural fragility has a median PR of 0.82 (boxplot summary = 0.68, 0.95, 0.81, 0.88; min, max, first quartile, third quartile). (c) A summary of the Cohen’s D effect size measurements between the success and failed outcome distributions across all features. The effect size of neural fragility is significantly greater then that of the beta band (alpha = 0.05). Neural fragility has a median effect size of 1.51 (boxplot summary = 0.92, 2.21, 1.43, 1.60; min, max, first quartile, third quartile). (d) The corresponding PValues of the effect size differences between success and failed outcomes, computed via the one-sided Mann-Whitney U-test. Note that the samples in (c) are not shown for visual sake; data was approximately bell-curve shaped and a box plot adequately summarizes the main descriptive statistics of the distribution. For box plot summary statistics (min, max, median, first quartile and third quartile) of other features, see Supplemental Data Table for this figure.