EEG epilepsy seizure prediction: the post-processing stage as a chronology

Almost one-third of epileptic patients fail to achieve seizure control through anti-epileptic drug administration. In the scarcity of completely controlling a patient’s epilepsy, seizure prediction plays a significant role in clinical management and providing new therapeutic options such as warning or intervention devices. Seizure prediction algorithms aim to identify the preictal period that Electroencephalogram (EEG) signals can capture. However, this period is associated with substantial heterogeneity, varying among patients or even between seizures from the same patient. The present work proposes a patient-specific seizure prediction algorithm using post-processing techniques to explore the existence of a set of chronological events of brain activity that precedes epileptic seizures. The study was conducted with 37 patients with Temporal Lobe Epilepsy (TLE) from the EPILEPSIAE database. The designed methodology combines univariate linear features with a classifier based on Support Vector Machines (SVM) and two post-processing techniques to handle pre-seizure temporality in an easily explainable way, employing knowledge from network theory. In the Chronological Firing Power approach, we considered the preictal as a sequence of three brain activity events separated in time. In the Cumulative Firing Power approach, we assumed the preictal period as a sequence of three overlapping events. These methodologies were compared with a control approach based on the typical machine learning pipeline. We considered a Seizure Prediction horizon (SPH) of 5 mins and analyzed several values for the Seizure Occurrence Period (SOP) duration, between 10 and 55 mins. Our results showed that the Cumulative Firing Power approach may improve the seizure prediction performance. This new strategy performed above chance for 62% of patients, whereas the control approach only validated 49% of its models.

▷ optimal parameters combination considering the highest final performance

Features description
Here are present some details regarding the extracted features.Linear features are mathematical measures that capture linear dynamics from the signal, using its phase/frequency and amplitude information.The EEG signal is assumed as quasi-stationary within each time window when this type of feature is extracted.

Statistical Moments
Statistical moments are widely used in seizure prediction studies to characterize the signal's amplitude distribution.The four moments are mean, variance, skewness, which measures the degree of asymmetries of the amplitude distribution, and kurtosis, which measures the relative flatness or peakedness of the amplitude distribution.The preictal period has been associated with considerable changes in these measures compared to the interictal period.In particular, a decrease in variance and an increase in kurtosis were observed in the preictal phase [2][3][4][5][6] .

Hjörth Parameters
The Hjörth parameters consider standard deviations to quantify the dynamical signal properties.These are three timedomain measures of brain activity: activity, a measure of mean power, mobility, a measure of root-mean-squared frequency, and complexity, a measure of root-mean-square frequency spread.With the proximity to the seizure onset, an increase in mobility and complexity measures is observed 2, 3, 5-7 .

Decorrelation Time
The decorrelation time is described as the first zero crossing of the autocorrelation function.It is an estimator of the data periodicity and the strength of linear correlations.The lower its values, the less the signal is correlated.Before seizures, a decrease in the decorrelation time has been reported 3,5 .

Relative Spectral Power
The spectral power quantifies the signal power associated with specific frequency ranges.It is possible to compute the power spectral density (PSD) by applying the Fast Fourier Transform (FFT) to the EEG time series and then average the squared coefficients of the frequency range of interest.
In turn, the relative spectral power is characterized as the power of a given frequency band divided by the total power of the EEG signal.A normalized spectral power provides a more robust measure since there is more power in low frequencies than at high frequencies.Some authors have reported a transference of power from the lower to higher frequencies before the seizure onset 3,[5][6][7] .

Spectral Edge Frequency and Power
SEF (Spectral Edge Frequency) is commonly described as the minimum frequency below which a given percentage of the total power of the signal is contained.The SEP (Spectral Edge Power) is the value of the power existing below the defined threshold.
Regarding the EEG signal, most of the spectral power is comprised in the 0.5-40Hz band, and SEF 50 and SEP 50 are commonly used.SEF 50 is the frequency below which 50% of the total power of the signal up to 40 Hz is located, and SEP 50 is the corresponding power below the spectral edge frequency.Thus, SEF may be capable of capturing the dynamics mentioned above during the preictal 3,5 .

Wavelet Coefficients Energy
The DWT (Discrete Wavelet Transform) is a time-frequency domain transform that can be an alternative to the FFT.It is capable of revealing the spectral and temporal properties of the signal.The wavelet transform decomposes the signal in different resolution levels according to specific frequency components.The first decomposition levels are associated with higher frequencies, while the last levels represent the lower frequencies.After the signal decomposition, it is possible to compute discriminant measures from distinct frequency bands by applying the wavelet coefficients.The quantification of the energy in different frequency ranges is an example of a feature that can be obtained using the wavelet transform 3,5 .

Results
Table S2 contains the seizure prediction results obtained for each patient and approach, including the optimal SOP value, seizure sensitivities (SS), false prediction rate per hour (FPR/h), and the models performing above the chance level.Table S3 presents the statistical test results for each patient and approach, using the seizure-times surrogates method.Tables S5 and S4 include the multiple comparison results using the Tukey's Honest Significant (HSD) Test.Seizure sensitivity might be influenced by the number of tested seizures.For patients with only one seizure, the model's sensitivity is limited to 0 1, indicating whether the seizure is correctly predicted or not.However, for patients with multiple the model capable of achieving different SS Furthermore, for patients with an elevated number of testing it might be more difficult to predict all achieve a seizure sensitivity of 1.
The following figures present the interaction between SS values and the number of testing seizures.Upon inspection, it is evident that the models achieved an SS value of 1 when the number of tested seizures was reduced.Moreover, more patients accomplished an SS value of 1 when only one seizure was tested.A general decrease in seizure sensitivity is observed as more seizures are tested.
Fig.S1.Interaction between seizure sensitivity (SS) and the number of testing seizures for each approach.Near to each point is the number of models that it represents.

Table S2 :
Seizure prediction results obtained for each patient and approach.

Table S3 :
Results obtained with the statistical using the seizure-times surrogates method.

Table S4 :
Pairwise comparison results for SS values using the multiple comparison Tukey HSD test.

Table S5 :
Pairwise comparison results for FPR/h values using the multiple comparison Tukey HSD test.The influence of the number of seizures in the models'