Abstracts

Rationale: Spatial characterization of the seizure onset zone (SOZ) or its underlying network is critical to the creation of a surgical plan for the treatment of epilepsy. An anatomical representation of the SOZ is created from both non-invasive data, including MRI, semiology, and EEG, and invasive stereotactic EEG (sEEG). Planning electrode placement for invasive monitoring also inherently depends on a conceptual model; typically, team members assess probable areas of seizure-onset based on prior results. Therefore, sEEG planning is an exercise in Bayesian optimization and information theory, with a goal of placing sEEG contacts in areas of uncertainty, requiring resolution before making a final surgical plan. We demonstrate the first stage of a Bayesian framework of SOZ localization in which non-invasive data can be represented as a 3D probability cloud, informing sEEG contact placement, and evaluates the information gained and overall utility of each sEEG contact placed in a retrospective study.

Methods: After retrospectively examining non-invasive data and the 3D coordinates of contacts placed in 20 sEEG patients, we created template SOZ probability clouds based on these prior data streams, as well as corresponding “key” SOZs. The probability clouds were described by the following values: lesioned, clear to diffuse margins, and single to multifocal. Lesion margin resolution is described with decay dependent on a varying Gaussian standard deviation. Multifocal cases modeled with graph theory vary in interconnectivity or causal outdegree between lesions. Further scripts simulate sEEG sampling strategies, including equidistant spacing, margin mapping, and single contact lesion sampling, emulating placement in the patients previously analyzed.

Results: We developed three main categories of sEEG strategy from the proposed models: defining margins along a variable Gaussian distribution, determining the causal target in a multifocal case, and broad sampling at points of discordancy. The 20 sEEG cases were qualitatively sorted into these categories by their principal problem, with 12 cases requiring margin definition, 3 multifocal cases, and 5 discordant or non-lesioned MRI. Many of these 20 cases could be described with a combination of these factors, leveraging the sliding scale of Gaussian distribution as well as varying reliance on outdegree or total interconnectivity.

Conclusions: This first attempt at an abstract model of sEEG placement for SOZ localization evidences a method of computationally planning invasive monitoring. The creation of quantifiable categories and corresponding sEEG placement strategies marks an important step forward in Bayesian inference in 3D sEEG contact planning. After this intuitive segregation, we can simulate localization using actual data from the 20 cases, building probability clouds and keys. Implanted sEEG contact informativeness can be graded by quantifying the information lost by omitting a sample point from localization calculations. The creation of these models now paves the way for the optimization of sEEG implantation.

Funding: Not applicable