Abstracts

Rationale: Stereoelectroencephalography (sEEG) is a surgical method that uses intracranial electrode recordings to identify brain networks participating in early seizure organization and propagation, referred to as the epileptogenic zone (EZ). Previously, we applied an unsupervised data-driven algorithm called Dynamic Mode Decomposition (DMD) to visualize, quantify and interpret sEEG data. DMD is a data reduction method able to generate vectors (called modes) and eigenvalues, which can be used to separate seizure recordings into frequencies, growth rates and spatial structures. We adapted DMD to produce Dynamic Modal Maps (DMMs) across frequency sub-bands, capturing epileptiform dynamics in the data and also providing a static estimate of EZ-localized contacts, termed the Higher-Frequency Mode-based Norm Index (MNI). Here, we look to improve the MNI by varying different settings---including modes with associated lower frequency combinations, high-to-low frequency ratios, different norm-related percentiles and epoch duration times, and also consider two statistical algorithms that detect abrupt changes in signals---which resulted in the creation of 10 different DMD-based indices.

Methods: We compare the performance of these indices as individual binary classifiers---evaluating their accuracy, sensitivity, specificity, positive and negative predictive values---by applying them to a set of 44 seizure recordings from a retrospective cohort of 10 patients previously diagnosed with temporal lobe epilepsy, who underwent sEEG implantation at our center. Of these 10 patients, 5 were treated by anteromesial temporal lobectomy (including resection of the amygdala and hippocampus), and 5 had a limited temporopolar resection (including the amygdala, but sparing the hippocampus). All 10 patients remained seizure-free for at least five years. We considered all resected anatomical electrode contacts as ‘ground truth.’ DMD-based indices were then generated for all seizures as separate events and also aggregated per patient, and results were compared across all patients, and also between the two surgical cohorts of patients.


Results: Considering the seizures either separately or per patient, the DMD-based indices with higher accuracy, sensitivity and specificity were those where a 90^th percentile of the norm-related modes was selected for the higher frequencies---comparable to our previous findings---but using a percentage of the seizure duration instead of fixed epoch intervals for the MNI calculation. Indices where either the two higher frequencies, or two lower frequencies, were combined with the change point algorithm were also highly accurate and specific to localize the EZ.

Conclusions: Certain DMD-based indices seem to offer high performance accuracy and specificity, and a difference can be observed among the two surgical groups based on some of the metrics used in this project. Our preliminary results support the use and integration of DMD to the visualization, quantification and interpretation of sEEG data, and highlight features that may be integrated into a machine learning classifier, to be used in surgical decision-making for patients with temporal lobe epilepsy.




Funding: None.