Authors :
Presenting Author: Caila Coyne, MEng – University of Alabama at Birmingham
Helen Brinyark, BS – University of Alabama at Birmingham
Rebekah Chatfield, BS – University of Alabama at Birmingham
Benjamin Cox, MD – University of Alabama at Birmingham
Arie Nakhmani, PhD – University of Alabama at Birmingham
Rachel Smith, PhD, MS, BS – University of Alabama at Birmingham
Rationale:
Intracranial EEG-derived dynamical network models (DNMs) represent the effective connectivity between channels and can capture both resting state (RS) and the response dynamics that are evoked during cortical electrical stimulation. The models are comprised of state transition matrices (𝑨) estimated using discrete 500 ms of iEEG data and the equation 𝒙[𝑡+1]=𝑨RS𝒙[𝑡] for RS and 𝒙[𝑡+1]=𝑨CCEP𝒙[𝑡]+𝑩u[𝑡] for cortico-cortical evoked potentials (CCEPs) where 𝒙[𝑡]∈ℝnx1 is the state vector describing the neuronal activity from each of the 𝑛 channels and the 𝑩u[𝑡] term accounts for the electrical stimulation input. Properties of these DNMs, such as impulse response estimation and the amount of network influence to and from each channel, have shown promise for seizure onset zone (SOZ) localization. However, the RS-DNMs are typically averaged over one 5–20-minute snapshot, and if single pulse electrical stimulation is performed, it often only occurs once during a patient’s stay. This leaves an open question of how varying brain states may affect the effective connectivity captured by the DNMs.
Methods:
To quantify the similarity in DNM connectivity between states, we calculate a transformation matrix, T , that maps one network model to another: 𝑨2 = T𝑨1. If the effective connectivity in both states is equivalent, T would be the identity matrix (I), so we calculate the transformation error (Terr) as RMSE(T – I). We can also assess the maximum singular value of T (Tσmax) as a proxy for transformation energy since σmax represents the maximum magnitude of deformation. We expect more similar brain states to have smaller Terr and Tσmax values.
Results:
We first assessed RS to RS transformations and found that both Terr and Tσmax are significantly lower for transformations between sequential DNMs compared to transformations between non-sequential brain states an average of 250 seconds apart (paired t-test p < < 0.01), indicating that the brain networks are modulating on the scale of seconds. We also show that a low number of randomized transformations are necessary for capturing the RS distribution as there were no significant differences in the metrics for 50 to 1000 transformation iterations (ANOVA p = 0.79). We then analyzed RS to CCEP transformations and found that 100% of stimulations of clinically annotated SOZ channels and 85.7% of early propagation zone channels resulted in significantly different transformations of the RS network compared to stimulation of non-SOZ channels (Wilcoxon p< < 0.01).