Brain-state Modeling for Adaptive Closed-loop Neuromodulation of Epilepsy
Abstract number :
1.295
Submission category :
3. Neurophysiology / 3G. Computational Analysis & Modeling of EEG
Year :
2024
Submission ID :
798
Source :
www.aesnet.org
Presentation date :
12/7/2024 12:00:00 AM
Published date :
Authors :
Presenting Author: Graham Johnson, MD, PhD – Mayo Clinic
Derek Doss, BE – Vanderbilt University
Ghassan Makhoul, BA – Vanderbilt University
Leon Cai, MD, PhD – Vanderbilt University
Camden Bibro, BS – Vanderbilt Univsersity
Danik Paulo, MD, MSCI – VUMC
Shilpa Reddy, MD, MMHC – Vanderbilt University Medical Center
Robert Naftel, MD – VUMC
Kevin Haas, MD, PhD – Vanderbilt University Medical Center
Mark Wallace, PhD – Vanderbilt University
Benoit Dawant, PhD – Department of Biomedical Engineering, Vanderbilt University
Angela Crudele, MD – Vanderbilt University Medical Center
Victoria Morgan, PhD – Vanderbilt University Medical Center
Christos Constantinidis, PhD – Vanderbillt University
Shawniqua Williams Roberson, MD – VUMC
Sarah Bick, MD – VUMC
Dario Englot, MD, PhD – Vanderbilt University Medical Center
Rationale: The progress of developing an effective closed-loop neuromodulation system for many neurological pathologies is hindered by the difficulties in accurately capturing a useful representation of a brain’s instantaneous functional state. Existing approaches rely on expert labeling of electroencephalography data to develop biomarkers of neurophysiological pathology. These techniques do not capture the highly complex functional states of the brain that are presumed to exist between labeled states or allow for the likely possibility of variation among identically labeled states. Thus, we propose BrainState, a self-supervised technique to model an arbitrarily complex instantaneous functional state of a brain using neural multivariate timeseries data.
Methods: To develop and validate the patient-specific brain-state model architecture, we utilized our cohort of approximately 17,000 hours of continuous SEEG data from 118 patients with drug-resistant epilepsy undergoing SEEG presurgical evaluation (Figure 1A-H). Next, we evaluated if the trained 1024-dimensional brain-state model effectively organized based on post-hoc inclusion of known peri-ictal labels by reducing the dimensionality to two and clustering the data – examples in Figure 2A-F. Finally, we tested the hypothesis that the brain-states could be selectively neuromodulated by single-pulse electrical stimulation (SPES).
Results: The projection of all SEEG data into a two-dimensional representation of the 1024-dimensional brain-state space allowed for clear self-organization of pre-ictal, ictal, and post-ictal epochs based on post-hoc inclusion of known peri-ictal labels (Figure 2G-J), including on withheld data. Neuromodulation of the state-space through SPES reveals increased brain-state transitions (t-test p-value range: 0.0367 to 6.25e-5) and increase in unique brain-states (p-values 8.81e-3 to 6.74e-8) during low-energy stimulation.
Conclusions: We have developed a self-organized patient-specific electrographic model of seizure propensity with evidence of neuromodulation during low-energy stimulation. We anticipate that BrainState will enable the development of sophisticated closed-loop neuromodulation systems for a diverse array of neurological pathologies.
Funding: F31NS120401
Neurophysiology