Table 1.
Summary of iEEG dataset and seizure onset.
Fig 1.
Power spectra of 1-min interictal (left) and preictal (right) epochs from 18 iEEG channels (patient 15).
Power spectrum in red is the average power spectrum of all channels. Both epochs show a power law behavior (P(f) ∼ 1/fα) illustrated with the fitted dotted red line for the range of frequencies delimited with a dashed blue line (corresponding to the range of scales j between 3 and 7). The exponents α obtained by a least-square fit are different in the preictal and interictal epochs.
Fig 2.
Estimation of cumulants c1 and c2 using wavelet leader and bootstrap based scaling analysis in signals with different scale invariance properties.
Each signal is 4096 samples. 100 bootstrap wavelet leader resamples were used in cumulant estimation. (a) Realization of a self-similar signal (fractional Brownian motion). ζ(q) is linear in q. c1 ≠ 0 and c2 ≈ 0. (b) Realization of a multifractal random walk. ζ(q) is nonlinear in q. c1, c2 ≠ 0. (c) EEG channel recording showing nonlinear relation of ζ(q) in q. c1, c2 ≠ 0. (Each column, top to bottom: Signal plot, regression plot of ζ(q) exponent estimates and boxplot of cumulant estimates).
Fig 3.
Flow chart of the seizure prediction method.
Steps 1 to 3 are training procedures and 4 to 6 are testing procedures. Boxes in orange show procedures used instead when combination of features are used.
Table 2.
Significance of the difference in the average cumulant and the spectral power observations between preictal and interictal epochs (5-min length) using the most discriminating channel in c1 and the most discriminating channel in c2.
Fig 4.
Box-and-whisker plot (minimum-maximum range) indicating differences in three patients (P2, P10 and P17) between preictal and interictal average 5-min observations of cumulant c1, cumulant c2 and the spectral power in the conventional EEG bands. Boxes in orange represent observations from the most discriminating channel in cumulant c1. Boxes in blue represent observations from the most discriminating channel in cumulant c2. Significant differences between preictal and interictal observations are denoted by asterisks (* p-value < 0.01, ** p-value < 0.001).
Fig 5.
Performance of the seizure prediction method using cumulants and their combination.
a-c top: Performance values for the range of persistence-τ values analyzed. Orange circles indicate interpolated values of sensitivity, proportion of time under warning and warning rate corresponding to the critical false prediction rate of 0.15/h. a-c bottom: Number of patients in whom seizures are predicted above chance as a function of the persistence-τ parameter. d. Average prediction time per patient as a function of the persistence-τ parameter.
Fig 6.
Performance of the seizure prediction method using combination of cumulants and state similarity measures.
a-c top: Performance metric values for the range of persistence-τ values analyzed when state similarity measures are combined with respectively cumulant c1 (feature set FS4), cumulant c2 (feature set FS5) and both cumulants (feature set FS6). Orange circles indicate interpolated values of sensitivity, proportion of time under warning and warning rate corresponding to the critical false prediction rate of 0.15/h. a-c bottom: Number of patients in whom seizures are predicted above chance level as function of the persistence-τ parameter, for each of the combination feature sets. d. Average prediction time per patient as a function of the persistence-τ parameter for each combination feature set.
Table 3.
Seizure prediction performance at the critical false prediction rate.