Automated classification of seizure onset pattern using intracranial electroencephalogram signal of non-human primates.

Objective: To develop and validate a machine learning framework for the classification of distinct seizure onset patterns using intracranial EEG (iEEG) recordings in a non-human primate (NHP) model of penicillin-induced seizures. Approach: iEEG data were collected from six NHPs, comprising 1,496 frontal and 549 temporal lobe seizures. Seizure onset patterns were manually categorized into five types: Sharp Activity (5-15 Hz), Low Amplitude Fast Activity (15-30 Hz), Delta Brush (1-3 Hz with bursts), High Amplitude Spike (2-5 Hz), and Polyspike. A Random Forest classifier was trained using features extracted from optimized seizure onset segments. Feature selection and seizure segment length optimization were performed using nested cross-validation to enhance classification accuracy and generalizability. Main results: The classifier achieved strong performance with F1-scores exceeding 79% for Sharp Activity, Low Amplitude Fast Activity, and High Amplitude Spike patterns. When validated on an independent temporal lobe seizure dataset, the model demonstrated robust generalizability, achieving precision and sensitivity above 80% for Sharp Activity and High Amplitude Spike. Significance: These findings demonstrate that the suggested spectral and dynamic features can effectively distinguish seizure onset patterns and generalize in distinct brain regions. Although there are limitations due to use of manual annotations and the sample size of certain categories, the proposed approach provides a framework for automatic classification of seizure onset patterns. Further, the framework has a potential use for epilepsy research and clinical applications in future.