Machine Learning Based Localization of the Epileptogenic Zone Using Cortico-cortical Evoked Potential Data
Abstract number :
2.067
Submission category :
3. Neurophysiology / 3G. Computational Analysis & Modeling of EEG
Year :
2022
Submission ID :
2204664
Source :
www.aesnet.org
Presentation date :
12/4/2022 12:00:00 PM
Published date :
Nov 22, 2022, 05:26 AM
Authors :
Sridevi Sarma, PhD – Johns Hopkins University; Shreya Singh, - – Johns Hopkins University; Dongwoo Lee, - – Johns Hopkins University; Mark Hays, BS – Johns Hopkins University; Rachel June Smith, PhD – Johns Hopkins University; Nathan Crone, MD – Johns Hopkins University; Joon Kang, MD – Johns Hopkins University
Rationale: Approximately 30% of patients have medically refractory epilepsy (MRE), where their seizures cannot be completely controlled by anti-epileptic drugs. A potential cure for MRE is the surgical removal of the epileptogenic zone (EZ), but this requires accurate identification of the brain regions that initiate and propagate seizure activity. Identification of the EZ remains an unsolved problem, thus surgical success rates currently range from 30%-70%. To improve localization of the EZ, we utilized machine learning (ML) methods to develop a predictive algorithm to identify epileptogenic brain regions, leveraging features from cortico-cortical evoked potentials (CCEPs) from single-pulse electrical stimulation (SPES).
Methods: We trained a random forest model on nodal and network features that were derived from CCEP data obtained from 19 patients who underwent intracranial EEG (iEEG) monitoring and SPES at Johns Hopkins Hospital. All patients underwent surgery and achieved at least 1 year of seizure freedom (Engel class I). Each iEEG contact was considered a node in the CCEP network that was influenced by the network when neighboring nodes were stimulated, and influential to the network when the node itself was stimulated. Features quantifying influence by and influential to the node were derived from CCEP peak amplitudes. The node’s importance within the CCEP network was captured by graph theory centrality measures, including eigenvector and closeness centrality. These nodal features were split into 75:25 train-test sets, and a random forest model was then trained for binary classification of whether an iEEG contact was within or outside the EZ. We evaluated the model on the test sets and assessed accuracy, sensitivity, and specificity alongside a receiver operator curve. Additionally, patients with failed surgical outcomes were analyzed to indicate EZ regions potentially misidentified by conventional clinical analysis.
Results: Our random forest model was able to distinguish EZ sites with an AUC of 87%, accuracy of 78%, sensitivity of 92%, and specificity of 75%. We found that the inclusion of nodal and network features predicted EZ locations with better accuracy than a model trained on either feature type alone. Further, the model often predicted locations for the EZ outside the resected regions for patients with failed surgical outcomes.
Conclusions: Our study pilots using a machine learning framework within CCEP data to identify EZ locations for surgical resection. This ML algorithm may serve as a complementary tool in the clinical workflow to suggest possible areas for resection, improving the chances of successful surgical outcomes.
Funding: Funding came from NIH IRACDA through JHU’s ASPIRE Program (RJS) and the Institute for Computational Medicine.
Neurophysiology