Cortico-cortical evoked potentials: Automated localization and classification of early and late responses

Cortico-cortical evoked potentials (CCEPs), elicited via single-pulse electrical stimulation, are used to map brain networks. These responses comprise early (N1) and late (N2) components, which reflect direct and indirect cortical connectivity. Reliable identification of these components remains difficult due to substantial variability in amplitude, phase, and timing. Traditional statistical methods often struggle to localize N1 and N2 peaks under such conditions.

A deep learning framework based on You Only Look Once (YOLO v10) was developed. Each CCEP epoch was converted into a two-dimensional image using Matplotlib and subsequently analyzed by the YOLO model to localize and classify N1 and N2 components. Detected image coordinates were mapped back to corresponding time-series indices for clinical interpretation.

The framework was trained and validated on intracranial EEG data from 9 patients with drug-resistant epilepsy (DRE) at the University of Alabama at Birmingham (UAB), achieving a mean average precision (mAP) of 0.928 at an Intersection over Union (IoU) threshold of 0.5 on the test dataset. Generalizability was assessed on more than 4000 unannotated epochs obtained from 5 additional UAB patients and 10 patients at the Hospital of the University of Pennsylvania.

To our knowledge, no existing deep learning methods localize and classify both N1 and N2 components, limiting comparison. Current approaches rely on manual identification within fixed windows, introducing inter-rater variability and often missing inter-individual differences.

The proposed framework accurately detects and classifies CCEP components, offering a robust, automated alternative to manual analysis.