Abstracts

Rationale: Electroencephalographic source localization (ESL) is a valuable component of presurgical evaluations for intractable epilepsy, particularly when resective surgery is considered in MR-negative cases. Clinical protocols for localizing the neuroanatomical ‘sources’ of epileptiform spiking and seizure patterns generally comprise modules for EEG signal processing, structural modeling, and rational constraints, controls, and visualizations. Similar modules also compose typical workflows for biophysical simulations, with numerous examples from the computational neurosciences over the past three decades. The "brute force" computational strategy for analyzing neural data entails using large numbers of model parameter combinations to map their relationships to model behavior. ESL protocols can be viewed as similar simulations with equivalent signal processing and anatomical parameters. Optimization of these parameters for each patient will improve the insights from EEG data for presurgical decision making. In the present work we adapt strengths and lessons-learned from BF computation to improve the quality of and confidence in EEG analytics and ESL.

Methods: As a quality improvement project we are adapting well-established computational modeling techniques to seizure source localization and computer-aided analyses of EEG. The major challenge to this initiative is to conduct appropriate BF analyses within the timeline of the clinical workflow during phase 1 admissions. In the first stage of this QI project we have built the core Matlab code architecture (MathWorks, Natick, MA) for analysis of hour-long epochs of EEG to extract timepoints of interest for ESL (Curry 9.0).

Results: First goals accomplished are: implementation of algorithms to (1) categorize subsets of stereotyped epileptiform activity patterns, (2) provide metrics of functional connectivity to evaluate the impact of distinct spike types and spread patterns on circuit dynamics. With these tools we begin to build circuit models of epileptogenic brain areas for each patient. (3) We also applied similar clustering algorithms to distinguish epileptiform events that should be grouped or excluded to improve upon standard ESL protocols by removing “non-linear averaging” across different event types even when they may look similar to the eye.

Conclusions: Using patients’ own imaging to build structural brain models is already part of most ESL protocols. To this we add patient-specific tuning of analysis and model parameters in order to improve accuracy, scope, and reliability of ESL contributions to presurgical evaluations

Funding: none