DeepSOZ: A Robust Deep Model for Joint Temporal and Spatial Seizure Onset Localization from Multichannel EEG Data

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Epilepsy diagnosis relies heavily on interpreting electroencephalogram (EEG) data.
• Accurate seizure detection and localization are crucial for effective treatment and surgical planning.
• Current methods may lack the precision and automation needed for seamless clinical integration.

Purpose of the Study:

• To introduce DeepSOZ, a novel deep learning framework for simultaneous seizure detection and localization from multichannel scalp EEG.
• To evaluate the robustness and performance of DeepSOZ compared to existing approaches.
• To assess the trustworthiness of DeepSOZ for clinical application in epilepsy management.

Main Methods:

• A transformer encoder generates global and channel-wise EEG encodings.
• A Long Short-Term Memory (LSTM) network performs temporal seizure detection.
• Attention-weighted multi-instance pooling predicts the seizure onset zone.
• The model was trained supervisedly and validated on a large dataset of 120 patients.

Main Results:

• DeepSOZ achieved robust performance in a multi-task learning setup.
• The framework provides high-resolution temporal (per second) and spatial (per EEG channel) predictions.
• Initial evaluations indicate good intra-seizure and intra-patient consistency.

Conclusions:

• DeepSOZ offers a robust and accurate deep learning solution for seizure detection and localization.
• The model's high-resolution outputs and demonstrated consistency are promising for clinical workflow integration.
• Further validation is recommended to establish DeepSOZ as a trustworthy tool for epilepsy care.