Rationale:
Predicting seizures from neural activity has long been of clinical interest, and computational algorithms for accurate prediction of seizure onset have recently received much attention. Research has generally focused on the balance between excitation and inhibition at the seizure focus. However, epilepsy is increasingly recognized as a network disorder. With this in mind, we hypothesized that the most predictive pre-seizure changes in activity may be distributed across cortical networks.
Method:
We sought to understand changes in activity patterns prior to seizure onset. We studied multisite spatiotemporal data from 24 hour intracranial electroencephalogram (iEEG) recordings in patients (London Health Sciences Centre, Western University) with therapy resistant epilepsy. We seek to understand patterns of narrow- and broad-band spectral shifts prior to seizure onset. We quantified changes in the 15-30 Hz β frequency band with respect to the wideband spectrum (1-100 Hz). By normalizing narrowband power by the wideband spectrum, which can account for differences in noise on individual electrodes, this technique can detect spectral shifts across electrodes while excluding potential confounds from analyzing narrow frequency bands in isolation.
Results:
Based on detected pre-seizure shifts in β, we developed a regularized classification model to identify spectral shifts that are predictive of seizure onset. We find predictions of seizure occurrence with high accuracy several minutes prior to seizure onset, with AUC metrics comparable to or surpassing previous approaches. Moreover, applying the classification model over the stream of iEEG signals reveals the abnormal brain activity during seizure vs non-seizure days, opening the possibility to provide advanced warning on days with higher risk of seizure occurrence. We then use a method for estimating network interactions from timeseries based on the precision matrix to understand the network of brain regions exhibiting predictive spectral shifts prior to seizure onset.
Conclusion:
By carefully taking subtle spectral shifts across distributed brain networks into account, we can predict seizure occurrence with high accuracy. Future work will apply these algorithms to a large database of iEEG recordings at LHSC-Western University. In addition, we will investigate how these distributed patterns underlie shifts in the balance of excitation and inhibition at the seizure focus through further computational analyses and models of cortical dynamics. These results provide insight into the significant but subtle differences in cortical network activity that precede seizure onset.
Funding:
:N/A
FIGURES
Figure 1