Fig. 5: Forecasting performance for 14 patients.
From: Critical slowing down as a biomarker for seizure susceptibility
a The pseudoprospective approach was applied to all patients, except Patient 5 due to too few seizures. A moving average (length 5 seizures) is shown for each patient. The number inset in each subplot represents the average risk level assigned to seizures. b The proportions of seizures correctly classified as high risk or incorrectly classified as low risk are compared for Method M1 (diamond), Method M2 (circle), and a random predictor (square). There was a significant effect of method used to forecast seizures in the high-risk category (F2,12 = 112.3, p = 7 × 10−13). A post-hoc analysis showed that both Methods M1 and M2 correctly classified significantly more seizures than the chance model (p = 9.6 × 10−10 and p = 9.8 × 10−10 for Methods M1 and M2, respectively). There were no significant differences between Methods M1 and M2 (p = 0.06). Methods M1 and M2 had significantly fewer seizures in the low-risk category than the random predictor (F2,12 = 148.3, p = 9.6 × 10−10 and p = 9.6 × 10−10 for Methods M1 and M2, respectively). There were no significant differences between Methods M1 and M2 (p = 0.5). c The proportions of time spent in the high and low risk categories for the three methods. No significant differences were found between the models (F2,12 = 1.13, p = 0.3 and F2,12 = 2.12, p = 0.1 for high and low risk categories, respectively). d Based on the performance product, there was a significant effect of method (F2,12 = 73.7, p = 6 × 10−11) and both Methods M1 and M2 performed significantly better than the chance predictor (p = 1.0 × 10−9 and p = 1.1 × 10−9 for Methods M1 and M2, respectively). There was no significant difference between Methods M1 and M2 (p = 0.2). Black symbols in plots b–d represent population means and lines represent ±one standard deviation. Gray symbols represent raw values for each patient (N = 13 patients). Statistical comparisons were computed using a balanced two-way ANOVA corrected with a Tukey–Kramer multiple comparisons test.
