A Machine Learning Paradigm to Identify Salient Brain Network Motifs in Drug-Resistant Epilepsy

Saturday, May 4, 2024

3:39 PM – 3:41 PM CT

Location: S405b

Presenting Author(s)

Derek J. Doss, BE

Medical Student
Vanderbilt University
Nashville, Tennessee, United States

Co-Author(s)

CB

Camden Bibro, BA

Research Assistant
Vanderbilt University, United States
Sarah K. Bick, MD (she/her/hers)

Assistant Professor of Neurological Surgery
Vanderbilt University Medical Center
Nashville, Tennessee, United States
CC

Catie Chang, PhD

Professor
Vanderbilt University, United States
Dario J. Englot, MD, PhD

Associate Professor of Neurological Surgery
Vanderbilt University Medical Center
Nashville, Tennessee, United States
MG

Martin Gallagher, MD, PhD

Neurologist
Vanderbilt University Medical Center, United States
Graham W. Johnson, PhD

Medical Student
Vanderbilt University
Antioch, TN, United States
GM

Ghassan Makhoul, BS

Medical Student
Vanderbilt University, United States
VM

Victoria Morgan, Vanderbilt University

Professor
PhD, United States
Robert Naftel, MD

Neurosurgeon
Vanderbilt University Medical Center
Nashville, TN, United States
AN

Aarushi Negi, BA

Undergraduate Student
Vanderbilt University, United States
SR

Shilpa Reddy, MD

Neurologist
Vanderbilt University Medical Center, United States
RR

Rohan Rashingkar, n/a

Undergraduate Student
Vanderbilt University, United States

Co-Author(s)

JS

Jared Shless, n/a

Medical Student
Harvard Medical School
San Francisco, California, United States

Co-Author(s)

SW

Shawniqua W. Williams Roberson, MD

Fellow
Vanderbilt University, United States

Introduction: Localization of seizure onset zones (SOZs) is necessary for surgical treatment of drug-resistant epilepsy (DRE). Network-based analyses of stereo-EEG (SEEG) have been used to localize SOZs, but the network signatures of SOZs are complex. Thus, we used a machine-learning method to characterize patient-specific network signatures that may localize SOZs.

Methods: 20-minutes of resting state SEEG data were obtained from 81 DRE patients. Each contact was classified as within a SOZ, propagative zone (PZ), or non-involved zone (NIZ). For each patient, 20 minutes of resting-state SEEG data was split into 2-minute epochs, where we calculated inwards and outwards alpha band [8-12 Hz] partial directed coherence (PDC). Principal component analysis was performed on each patient’s PDC matrix. Each principal component was ranked, and coefficients and weights were analyzed for connectivity in specific lobes and SOZ designations. To analyze the influence of one component on each patient's connectivity, weights, and coefficients were used to reconstruct original data.

Results: Despite PCA being run independently for each patient, a theme of high inward and low outward connectivity to SOZs was observed. This suppressive signature was the first principal component for 64/81 (79%) patients and exemplified higher explained variance than other components (p < 0.001, paired T. The remaining 17 patients still had a component with a suppressive signature, but it was not the first component. Both PC 1 and PC 2+ sub-cohorts have sequentially tiered separation between SOZ, PZ, and NIZ, exemplifying typical suppressive behavior (p < 0.001, one-way ANOVA). When reconstructing PDC data from the respective suppressive component, this separation remains (p < 0.001, one-way ANOVA).

Conclusion : Upon examination of the top components for each patient, we noticed a theme in-line with the suppressive component despite PCA being run on each patient. However, connectivity analyses of reconstructed data reveal underlying phenomena in the PC 2+ cohort governing the separation of SOZ, PZ, and NIZ. To properly understand the nature of this sub-cohort and their candidacy for surgery, more should be done to identify these phenomena.