SeizureID(TM): Accelerating Physician Review of Intracranial EEG Recordings from the RNS® System using AI
Abstract number :
1.253
Submission category :
3. Neurophysiology / 3E. Brain Stimulation
Year :
2025
Submission ID :
872
Source :
www.aesnet.org
Presentation date :
12/6/2025 12:00:00 AM
Published date :
Authors :
Presenting Author: Muhammad Furqan Afzal, PhD – NeuroPace, Inc.
Sharanya Arcot Desai, PhD – Samsung Research America
Wade Barry, BA – NeuroPace, Inc.
Thomas Tcheng, PhD – NeuroPace, Inc.
Jonathan Kuo, MD – Keck School of Medicine of USC
Shawna Benard, MD – Keck School of Medicine of USC
Christopher Traner, MD – Cleveland Clinic
Jacob Norman, PhD – NeuroPace, Inc.
Jatin Sharma, MS – NeuroPace, Inc.
Brett Wingeier, PhD – NeuroPace, Inc.
David Greene, BS – NeuroPace, Inc.
Cairn Seale, MS – NeuroPace, Inc.
Martha Morrell, MD – NeuroPace
Rationale: Manually reviewing intracranial electroencephalography (iEEG) recordings is time-consuming and labor-intensive. Rapidly and accurately classifying large volumes of iEEG data into seizure and non-seizure activity is important for improving the efficiency of physician review of iEEG data.
Methods: Intracranial EEG recordings from focal epilepsy patients enrolled in clinical trials of the NeuroPace RNS® system were used to train and evaluate an AI model for classifying electrographic activity as seizure or non-seizure. Each recording typically contained 4 channels, with electrodes implanted in either the mesial temporal lobe or neocortex. A total of 136,878 iEEG recordings, representing approximately 550,000 channels of data, from 113 patients were used. The dataset was split into 72 patients for training, 18 for tuning, and 23 for testing, and was labeled by a NeuroPace employee. To accelerate the labeling process, a Bayesian Gaussian Mixture Model was used to cluster low-dimensional feature representations of iEEG channels generated by a convolutional neural network. Cluster centroids were manually labeled, and the assigned label was applied to all recordings within each cluster. An additional 100 patients were used in clinical validation, where 1000 iEEG recordings were independently labeled by three board-certified epileptologists.
Results: On the 23-patient test set, the vision transformer achieved a seizure classification accuracy of 96.7%. On the final clinical validation dataset from 100 patients, the model reached 96% accuracy and an F1-score of 95% on channels where all three experts agreed on the seizure vs. non-seizure labels. Performance was higher when using color spectrogram images compared to grayscale (96.7% vs. 95% accuracy, not statistically significant), highlighting the benefit of leveraging pretrained models designed for RGB inputs.
Conclusions: With 96% seizure classification accuracy on the clinical validation dataset, this tool has demonstrated potential clinical utility. The vision transformer model is under development for integration into the Patient Data Management System (PDMS) physician portal and may provide an effective tool for accelerating iEEG review to support efficient clinical decision-making.
Funding: Not applicable.
Neurophysiology