Authors :
Presenting Author: Alejandro Nieto Ramos, PhD – Cleveland Clinic Epilepsy Center
Balu Krishnan, PhD – Cleveland Clinic Epilepsy Center; Neha John, MSc – Cleveland Clinic Epilepsy Center; Juan Bulacio, MD – Cleveland Clinic Epilepsy Center; Demitre Serletis, MD PhD FRCSC FAES – Cleveland Clinic Epilepsy Center
Rationale:
Dynamical behavior, as observed in the cardiac or nervous systems, is characterized by high-dimensional data that is difficult to quantify. In the brain, the investigation of governing spatiotemporal dynamics underlying transitory "states" has become the focus of clinicians and scientists alike. Epilepsy, in particular, represents the epitome of dynamical disorders afflicting the brain. Importantly, patients undergoing stereoelectroencephalography (sEEG) offer a unique opportunity to directly record from brain networks. At present, analysis of sEEG data is limited to qualitative review by clinical experts. Given significant challenges with visualization and interpretation of this big data, we propose the use of data-driven methods to better define the nonlinear, multi-scale complexity intrinsic to epileptiform activity. Here, we employ a mathematical approach called "Dynamic Mode Decomposition" (DMD) to obtain linear representations of strongly nonlinear dynamics intrinsic to sEEG data, in order to generate spatiotemporal coherent structures called "modes" that define the dominant signal features. Modal evolution can be modeled over time, permitting the identification of unknown dynamics from the data itself. In this project, we employ DMD to describe sEEG dynamics as a function of spatially-correlated modes and develop tools for epileptic network identification and dynamical estimation.
Methods:
We analyzed representative recordings from three refractory epilepsy patients undergoing sEEG at the Cleveland Clinic, who underwent neurosurgical resection and have been seizure-free for a minimum of one year. EEG recordings were collected using a Nihon Kohden amplifier and sampled at 1 kHz. Data was band-pass filtered with cut-off frequencies of 0.1 and 300 Hz and converted to a bipolar montage using Brainstorm Matlab toolbox. DMD analysis was performed using Matlab R2022b.
Results:
For each representative case, DMD was able to capture clinically-meaningful data from the sEEG recordings. Notably, the method highlighted sEEG contacts capturing the most unique (least frequent) modes, revealing profile maps (i.e., dynamic modal maps) for the electrode/contact pairs most heavily implicated in early organization and spread of seizure activity. The method was successful in the analysis of both temporal lobe epilepsy cases and also a complex frontomesial epilepsy case. From these results, we developed a visualization tool using frequency sub-band analysis and phase locking metrics from the computed modes, suggesting that higher frequencies (gamma/beta bands) are more effective at network identification of early epileptiform activity.
Conclusions:
Data-driven methods offer mathematical insight into identification of epileptic brain networks and dynamical estimation of signal properties. Our preliminary results successfully identified low-dimensional biomarkers that may be used to better define the epileptogenic zone. Our proposed next steps will extract modal-specific dynamics out of a larger patient cohort incorporating machine learning workflows to analyze sEEG-based estimates of dynamical connectivity in epilepsy.
Funding: None.