PREDICTING EPILEPTIC SEIZURE RISK WITH SINGLE NEURON ACTIVITY DURING RESTING PROTOCOLS

For patients with drug-resistant epilepsy, reliable seizure warnings are essential. Predicting the proictal state, characterized by an increased likelihood of seizures, is crucial for enhancing safety and quality of life. Continuous recordings were obtained from 25 patients over a three‑week period using hybrid Behnke‑Fried macro‑micro electrodes implanted in mesial temporal lobe structures. Our analyses focused on the microelectrode intracranial EEG (iEEG) signals collected during standardized daily resting‑state protocols. After spike‑sorting, we identified 1,237 single units, which were subsequently classified according to their putative neuronal subtype, their mesial temporal region and wether they were localized within the seizure onset zone. Each single unit was labeled according to its temporal relation to the next seizure, being designated interictal if recorded during seizure-free days or proictal if recorded within the 24 hours preceding a seizure. We subsequently derived interpretable, domain-specific features capturing the behavioral dynamics and firing characteristics of these neurons. We initially evaluated the statistical significance of each feature using a mixed‑effects model that included mesial temporal region and the proictal or interictal label as variables. Subsequently, we applied a customized under-bagging strategy to train a range of machine learning classifiers, mitigating class imbalance while assessing the relative importance of each feature. We will present key findings from the statistical analyses and the assessment of predictive model performance. Together, these results advance our understanding of neuronal activity during the proictal transition and may inform the development of more accurate, localized seizure prediction models.