Authors :
Kira Kiviat, MS – University of California at Davis
Zachary McNaughton, BS – University of California at Davis
Jivesh Patel, BS candidate – University of California at Davis
Caren Armstrong, MD PhD – UC Davis
Presenting Author: Sheela Toprani, MD, PhD – University of California at Davis
Rationale:
Epilepsy patients can have multiple types of seizures, originating from different locations and following different spread patterns. It is important to identify these distinct types to understand where intervention may be required to treat the seizures. While these seizure types can be determined by visually inspecting intracranial EEG (iEEG) signals, it is useful to (1) computationally characterize seizure types (e.g., by effective connectivity) to determine whether visual differences reflect quantifiably distinct seizures versus different parts of a spread pattern, (2) be able to characterize future seizures to determine if they are the same as those already characterized or a new type, and (3) compare stimulated seizures to each typical seizure type or distinguish atypical/aberrant stimulated seizures. It is important for clinical decision-making to know whether a stimulated seizure is representative of typical spontaneous seizure to determine where to focus surgical treatment.
Methods:
We have characterized continuous iEEG data over 1-3 weeks during seizures (3-60 per patient) by calculating spectral coherence in a series of time windows between all pairs of recording electrodes in a range of physiological frequency bands to estimate effective connectivity across the brain network in patients with refractory epilepsy undergoing surgical planning. These pairwise coherence values define a location in a high-dimensional state space for each time window. Using the dimensionality reduction algorithm PaCMAP (Pairwise Controlled Manifold Approximation)[1], we can visualize this high-dimensional space in low dimensions while preserving important structure. This separation is made more rigorous by reducing the dimensionality of the data through principal component analysis and applying logistic regression to classify seizure types.Results:
This study assessed 5 different seizure types (N=100 seizures) across three patients. In patients (N=2) with multiple seizure types, PaCMAP analysis shows different seizure types forming separate effective connectivity clusters, while in a patient with only one seizure type (N=1), the seizures do not separate. Correspondingly, the distinct seizure types can be successfully classified using logistic regression. In a patient with stimulated seizures, both PaCMAP visualization and logistic regression showed that the stimulated seizures could not be distinguished from spontaneous seizures. This is consistent with the clinical evaluation that the stimulated seizures were typical.Conclusions:
Spectral coherence applied to iEEG contacts defines a state space that can separate distinct seizure types, and may be able to distinguish atypical stimulated seizures from typical ones.
Funding:
National Center for Advancing Translational Sciences, National Institutes of Health UL1 (# TR001860) and linked UC Davis Clinical Translational Science Center (CTSC) KL2 Award (# TR001859)