AI-Driven Electrographic Seizure Classification and Seizure Onset Detection Using Image- and Time-Series-Based...

Area of Science:

• Artificial Intelligence in Neurology
• Computational Neuroscience
• Medical Device Technology

Background:

• Manual analysis of intracranial electroencephalography (iEEG) for seizure detection is time-consuming.
• Accurate electrographic seizure classification (ESC) and seizure onset detection (SOD) are crucial for epilepsy management.
• Treatment-resistant epilepsy patients require efficient and precise diagnostic tools.

Purpose of the Study:

• To explore AI-based approaches for ESC and SOD in iEEG recordings.
• To evaluate the performance of various AI models, including Vision Transformers (ViTs), for seizure analysis.
• To assess the potential of AI to aid in personalized epilepsy treatment strategies.

Main Methods:

• Assessed image-based and time-series-based AI models (CNNs, ViTs, time-series transformers) for ESC and SOD.
• Utilized a large dataset of iEEG traces from 291 focal epilepsy patients with the NeuroPace RNS® system.
• Conducted extensive experiments varying model parameters, input formats, and initializations.

Main Results:

• Vision Transformers (ViTs) demonstrated superior performance in both ESC and SOD tasks.
• ViTs achieved 97% accuracy for ESC on cross-validation and 96% on a clinician-annotated test set.
• ViTs achieved a median absolute error (MAE) of 1.4s for SOD on cross-validation and 0.8s on the test set.

Conclusions:

• Image-based AI, particularly ViTs, effectively captures seizure patterns from iEEG data.
• AI-driven ESC and SOD can reduce manual review time for clinicians.
• AI tools hold promise for developing personalized epilepsy treatment strategies.