A Generalizable YOLO-Based Framework for iEEG Time Series Event Detection
Status: Work in Progress — Manuscript in Preparation Code, data, and model weights will be released upon publication.
This project adapts the YOLO (You Only Look Once) object detection framework from computer vision to one-dimensional intracranial EEG (iEEG) time series. The goal is automated, high-throughput, multi-class detection of seizure onset patterns (SOPs) directly from raw iEEG recordings — treating EEG pattern events as temporal objects with onset, offset, and class label.
Seizure onset patterns are clinically significant biomarkers in the evaluation of drug-resistant epilepsy patients undergoing surgical workup:
- They help localize the seizure onset zone (SOZ)
- Specific patterns (e.g., low-voltage fast activity) are associated with favorable surgical outcomes
- Manual annotation by experts is time-consuming and does not scale to large datasets
Existing automated approaches struggle with the variability inherent in SOPs: events span vastly different time scales (sub-second to tens of seconds) and fall into multiple distinct pattern classes.
We reformulate EEG pattern detection as a temporal object detection problem. A 1D implementation of the YOLO architecture processes multi-channel iEEG signals and simultaneously predicts temporal bounding boxes and class labels for multiple pattern types.
The model follows the standard YOLO pipeline adapted to the temporal domain:
| Component | Role |
|---|---|
| Backbone | Feature extraction from 1D iEEG signal |
| Neck (FPN/PAN) | Multi-scale feature fusion |
| Head | Temporal bounding box regression + class prediction |
Expert-annotated iEEG recordings serve as ground truth, produced via a dedicated seizure annotation interface.
Zhongchuan Xu, Carlos A. Aguila, Odile Feys, Dan Zhou, Honda Wu, William K.S. Ojemann, Nishant Sinha, Erin C. Conrad, Brian Litt
Preprint and paper link to be added upon publication.
For inquiries about this work, please reach out via GitHub Issues once the repository is fully public.