Authors :
Chenda Duan, BS, MS – University of California, Los Angeles
Yipeng Zhang, MS, PhD – Department of Electrical and Computer Engineering, University of California Los Angeles
Sotaro Kanai, MD, PhD – Division of Pediatric Neurology, Department of Pediatrics, David Geffen School of Medicine at the University of California, Los Angeles, California, USA
Yuanyi Ding, MS – Department of Electrical and Computer Engineering, University of California Los Angeles
Atsuro Daida, MD, PhD – Saitama Children's Medical Center
Pengyue Yu, BS – University of California, Los Angeles
Tiancheng Zheng, BS – University of California, Los Angeles
Naoto Kuroda, MD, PhD – Wayne State University
Shaun A. Hussain, MD, MS – Division of Pediatric Neurology, Department of Pediatrics, UCLA Mattel Children's Hospital, David Geffen School of Medicine
Eishi Asano, MD, PhD – Wayne State University
Presenting Author: Hiroki Nariai, MD, PhD, MS – Department of Pediatrics, Division of Pediatric Neurology, David Geffen School of Medicine at the University of California, Los Angeles, California, USA
Vwani Roychowdhury, PhD – Department of Electrical and Computer Engineering, University of California Los Angeles
Rationale:
Recent advances in machine learning have highlighted the potential of Intracranial EEG (iEEG) analysis for epilepsy surgery planning, yet progress has been slowed by the limited availability of large datasets with standardized clinical labels. Many existing efforts rely on inhomogeneous datasets, and often lack unified, clinically meaningful tasks and evaluation protocols. Omni-iEEG addresses these challenges through a multi-institutional, pre-surgical iEEG dataset with harmonized metadata, expert-reviewed annotations, and standardized benchmarks aligned with clinical objectives such as localization of the epileptogenic zone and outcome prediction (Figure 1).Methods:
The Omni-iEEG dataset includes 178 hours of pre-surgical interictal iEEG recordings from 302 patients across eight epilepsy centers (Table 1A). Of those, 232 patients have seizure onset zone (SOZ) annotations and 233 patients have resection margins with postoperative seizure outcomes after one year. Expert epileptologists annotated over 36,000 high-frequency oscillation (HFO) events, first detected using standard algorithms (e.g., STE, MNI), and then labeled as artifacts, HFOs without spikes (physiological HFOs), or HFOs with spikes (pathological HFOs), forming an event classification task. For pathological region identification tasks, we adopt a clinically guided labeling strategy. Channels within the SOZ are treated as positive examples of pathological regions. Preserved channels from seizure-free patients are treated as negative examples, as the resection is assumed to have removed all epileptogenic tissue. We also define exploratory tasks including anatomical region classification (frontal, temporal, limbic, parietal, and occipital), ictal period detection, and sleep–awake classification. All tasks use standardized preprocessing pipelines, training utilities, and unified evaluation protocols.
Results:
We evaluate baseline machine learning models across two settings: event-based models, which classify candidate HFOs, and segment-based models, which operate directly on raw iEEG segments. Prior work has applied vision-based models like convolutional neural networks (CNN) on spectrograms; we extend this by benchmarking state-of-the-art CNN-based SEEG-NET (Wang et al., Comput Biol Med, 2022) and audio models, such as AST(Gong et al., Proc Interspeech, 2021) and CLAP (Wu et al., ICASSP, 2023). Table 1B reports performance of segment-based models. The results highlight both the clinical feasibility and modeling challenges of the proposed tasks. Notably, audio-pretrained models (CLAP) perform well even on exploratory tasks with limited labeled data, suggesting that iEEG signals may benefit from cross-domain transfer learning.Conclusions:
Omni-iEEG is a publicly released, BIDS-compliant dataset and benchmark designed to bridge the machine learning and clinical communities. It integrates expert-labeled biomarkers and standardized, clinically meaningful tasks to enable reproducible iEEG research. A full utility library is provided to support reproducible benchmarking and clinical research.Funding: the National Institute of Neurological Disorders and Stroke (NINDS) K23NS128318