Authors :
Presenting Author: Saarang Panchavati, BS – UCLA
Atsuro Daida, MD, PhD – Saitama Children's Medical Center
Sotaro Kanai, MD, PhD – Division of Pediatric Neurology, Department of Pediatrics, David Geffen School of Medicine at the University of California, Los Angeles, California, USA
Aria Fallah, MD, MSc, MBA – Department of Neurosurgery, UCLA Medical Center, David Geffen School of Medicine
Vwani Roychowdhury, PhD – Department of Electrical and Computer Engineering, University of California Los Angeles
William Speier, PhD – Department of Radiological Sciences and Bioengineering, University of California Los Angeles
Hiroki Nariai, MD, PhD, MS – Department of Pediatrics, Division of Pediatric Neurology, David Geffen School of Medicine at the University of California, Los Angeles, California, USA
Rationale:
Neuromodulation of the thalamus is an emerging treatment for drug-resistant epilepsy, while little is known about the ictal thalamocortical dynamics in human EEG data. We characterized ictal thalamocortical dynamics via intracranial EEG (iEEG) during focal seizures and investigated whether thalamic iEEG signals can predict seizure states from onset to termination.
Methods:
We retrospectively analyzed 66 seizures derived from 19 patients with pediatric-onset epilepsy admitted to UCLA Mattel Children’s Hospital between November 2020 and July 2024 who underwent chronic stereotactic EEG targeting the cortex and the thalamus, sampling from the anterior (AN) and/or centromedian (CM) nuclei. All patients had medication-refractory focal epilepsy and were considered for thalamic neuromodulation. For each seizure, we epoched two 30-second time windows of interest, centering the ictal onset and termination. SEEG data were segmented into non-overlapping 2-second windows, and spectral power and connectivity, utilizing imaginary coherence and spectral Granger causality, were computed across canonical frequency bands: slow band (1-12 Hz), beta band (12–30 Hz), and gamma band (30–70 Hz). These values were normalized to an interictal baseline (−20 to −19 min pre-onset), and statistical differences were assessed using Wilcoxon signed-rank tests. Additionally, we trained random forest classifiers using features derived from the thalamus and the thalamocortical network to predict seizure states. Model performance was evaluated using the area under the ROC curve (AUROC), computed per seizure and across all predictions. Additionally, we used SHAP (SHapley Additive exPlanations) values to estimate feature importance across models trained on thalamic and thalamocortical features.
Results:
At seizure onset, there was a significant increase in power at both the cortex and thalamus, accompanied by bidirectional thalamocortical connectivity, all of which markedly decreased at seizure termination. Notably, these changes were more prominent in the slow frequency bands within the thalamus and thalamocortical network, whereas higher-frequency activity was primarily observed in the cortex (Figure 1). Using features from the anterior nucleus (AN), seizure state was classified with an AUROC of 0.825 ± 0.162. The most important predictors included beta, gamma power, and slow-frequency outflow from the AN to the seizure onset zone (SOZ). Using centromedian nucleus (CM) features, classification achieved an AUROC of 0.840 ± 0.149, with key features including slow-frequency and beta spectral power, as well as slow-frequency outflow from CM to the SOZ. (Figure 2).
Conclusions:
The thalamocortical network was consistently activated throughout the ictal period. Increased spectral power and directed connectivity in both the fast and slow frequency bands reliably predicted each ictal state, suggesting a mechanistic role for these frequency bands in ictogenesis. These neurophysiological features may also serve as biomarkers to guide therapies such as neuromodulation.
Funding:
National Institute of Neurological Disorders and Stroke, K23NS128318, Uehara Memorial Foundation and the SENSHIN Medical Research Foundation