Organizations: UCLA Samueli School of Engineering · UCLA Mattel Children’s Hospital, David Geffen School of Medicine · Children’s Hospital of Michigan, Wayne State University School of Medicine
Epilepsy affects over 50 million people worldwide, and one-third of patients suffer drug-resistant seizures where surgery offers the best chance of seizure freedom. Accurate localization of the epileptogenic zone (EZ) relies on intracranial EEG (iEEG). Clinical workflows, however, remain constrained by labor-intensive manual review. At the same time, existing data-driven approaches are typically developed on single-center datasets that are inconsistent in format and metadata, lack standardized benchmarks, and rarely release pathological event annotations, creating barriers to reproducibility, cross-center validation, and clinical relevance. With extensive efforts to reconcile heterogeneous iEEG formats, metadata, and recordings across publicly available sources, we present Omni-iEEG, a large-scale, pre-surgical iEEG resource comprising 302 patients and 178 hours of high-resolution recordings. The dataset includes harmonized clinical metadata such as seizure onset zones, resections, and surgical outcomes, all validated by board-certified epileptologists. In addition, Omni-iEEG provides over 36K expert-validated annotations of pathological events, enabling robust biomarker studies. Omni-iEEG serves as a bridge between machine learning and epilepsy research. It defines clinically meaningful tasks with unified evaluation metrics grounded in clinical priors, enabling systematic evaluation of models in clinically relevant settings. Beyond benchmarking, we demonstrate the potential of end-to-end modeling on long iEEG segments and highlight the transferability of representations pretrained on non-neurophysiological domains. Together, these contributions establish Omni-iEEG as a foundation for reproducible, generalizable, and clinically translatable epilepsy research. The project page with dataset and code links is available at omni-ieeg.github.io/omni-ieeg.
Figures & tables
Figure 1: Overview of the Omni-iEEG Dataset and Benchmark.
Split
Dataset
#Subjects
#Selected
#Surgery
#Seizure-Free
Duration (Min)
Training
OpeniEEG
111
111
97
66
5-111
Zurich
12
12
9
9
1-8
HUP
42
9
9
6
5
SourceSink
21
20
12
10
1-16
Testing
OpeniEEG
74
74
67
48
5-120
Zurich
8
8
6
6
3-12
Table 1: Overview of the Omni-iEEG dataset. “#Subjects” indicates the total number of subjects in the dataset; “#Selected” denotes those included in the current split; “#Surgery” refers to subjects who underwent resective surgery; “#Seizure-Free” indicates subjects achieves post-operative seizue-free; “Duration” specifies the range of recording lengths in minutes.
Split
Dataset
#Subjects
#Total Events
#Artifact
#Real HFO
#spkHFO
Training
OpeniEEG
20
19909
5517
14392
10070
Zurich
3
829
242
587
449
HUP
3
178
42
136
120
SourceSink
3
1120
949
171
121
Testing
OpeniEEG
14
13092
2052
11040
7964
Zurich
2
443
248
195
102
Table 2: Summary of HFO event annotations. “#Subjects” denotes the number of patients; “#Total Events” refers to all annotated HFO events; “#Artifact” indicates events labeled as artifacts; “#Real HFO” represents real HFO; and “#spkHFO” refers to real HFOs that co-occur with spikes.
Split
Dataset
# Valid Subjects
# Normal Channels
# Pathological Channels
# Seizure Free
# Seizure Non-Free
# Anatomical
Training
OpeniEEG
110
4638
1072
66
32
111
Zurich
12
5619
—
9
3
—
HUP
9
1034
106
6
3
—
SourceSink
19
500
177
10
9
—
Testing
OpeniEEG
74
3533
575
48
19
74
Zurich
8
2611
—
6
2
—
Table 3: Summary of channel labels and clinical outcomes across dataset splits. “#Subjects” indicates the number of valid subjects; “#Normal/Pathological Channels” represent iEEG channels labeled as normal or pathological; “#Seizure-Free/Non-Free” correspond to subjects with or without post-surgical seizure freedom; “#Anatomical Labeled” lists subjects with anatomical channel annotations.
Model
Precision
Recall
F1
AUC
LSTM+Attention
0.7352
0.7359
0.7338
0.9109
PatchTST Transformer
0.7757
0.7686
0.7726
0.9311
TimesNet
0.7589
0.7726
0.7652
0.9221
PyHFO-Omni
0.8025
0.8110
0.8061
0.9390
Table 4: Benchmark for Pathological Event Classification. Evaluation metrics include macro-averaged precision, recall, F1 and AUC.
Pathological Channel Classification
Outcome
Model
Precision
Recall
F1
Specificity
AUC
AUC
Event-Based Models
eHFO
0.6053
0.6466
0.6195
0.4101
0.6611
0.4521
PyHFO spkHFO
0.6000
0.6431
0.6140
0.4089
0.6557
0.4972
PyHFO-Omni spkHFO
0.5799
0.6991
0.5635
0.6951
0.7351
0.7438
Segment-Based Models
Table 5: Benchmark results for identifying pathological brain regions. Evaluation has two dimensions: (1) Pathological Channel Identification , based on clinical priors of pathological(SOZ channels) vs. normal (preserved channels in seizure-free patients); (2) Outcome AUC , which assesses whether higher scores in resected channels correlate with postoperative seizure freedom via the resection ratio.
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
Value
filter_freq
[80, 300]
rms_window
3×10−3
min_window
6×10−3
min_gap
10×10−3
min_osc
6
rms_thres
5
Appendix
Table 6: Detector parameters used in our annotation pipeline. All frequency units are in Hz, time in seconds unless noted.
Annotator(s)
Metric
Value
R1 vs. R2
Cohen’s κ
0.9427
R3 vs. R1
Cohen’s κ
0.9243
R3 vs. R2
Cohen’s κ
0.9091
R1, R2, R3
Fleiss’ κ
0.9254
Appendix
Table 7: Inter-rater agreement among the three primary annotators. Pairwise agreement is measured using Cohen’s κ , and overall agreement is measured using Fleiss’ κ .
Model
Anatomical(5)
Anatomical(12) )
Ictal
Sleep–Awake
F1
Acc
F1
Acc
F1
Acc
F1
Acc
SEEG-NET Wang et al. (2022)
0.2520
0.2550
0.1081
0.1853
0.7526
0.7624
0.6773
0.6421
CLAP Wu et al. (2023b)
0.4750
0.4894
0.3540
0.3897
0.9245
0.9323
0.7225
0.9331
TimeConv-CNN
0.4788
0.4708
0.3087
0.3215
0.8533
0.8720
0.7118
0.9291
Appendix
Table 8: Macro-averaged F1 and balanced accuracy scores for the Anatomical with 5 Classes and 12 Classes, Ictal/Interictal, and Sleep/Awake classification tasks.
Figure 2: Event-based iEEG input: raw waveform (left) and time-frequency representation (right).
Figure 3: Channel Segment-based iEEG input: raw waveform (left) and time-frequency representation (right).
Held-Out
Precision
Recall
F1
Open-iEEG
0.6955
0.6893
0.6227
Zurich
0.7342
0.7520
0.7416
HUP
0.6967
0.7647
0.7219
SourceSink
0.7110
0.7412
0.7221
Appendix
Table 9: Cross-dataset generalization via leave-one-dataset-out HFO event classification.
Held-Out
Model
Precision
Recall
F1
Specificity
AUC
Outcome
Open-iEEG
TimeConv-CNN
0.5912
0.6393
0.5936
0.7383
0.6983
0.5325
CLAP
0.5785
0.6207
0.5781
0.7284
0.6560
0.5122
SEEGNET
0.5746
0.6335
0.5279
0.5608
0.6650
0.4628
Source-Sink
TimeConv-CNN
0.5995
0.5857
0.5902
0.8392
0.6031
0.7740
CLAP
0.6109
0.6063
0.6083
0.8168
0.6222
0.7592
SEEG-NET
0.6965
0.5999
0.6100
0.9430
0.6331
0.7740
Appendix
Table 10: Cross-dataset generalization via leave-one-dataset-out pathological brain region identification.
Model
Precision
Recall
F1
Specificity
AUC
TimeConv-CNN (30 s)
0.5768
0.7072
0.5442
0.6592
0.7729
TimeConv-CNN (1 min)
0.6081
0.7607
0.6096
0.7478
0.8229
TimeConv-CNN (2 mins)
0.5919
0.7473
0.5642
0.6684
0.8052
Appendix
Table 11: Segment length ablation study (pathological channel classification).
Model
Precision
Recall
F1
Specificity
AUC
Outcome
TimeConv-CNN
0.5924
0.7192
0.5884
0.7371
0.7735
0.7113
CLAP
0.5997
0.7191
0.6065
0.7740
0.7843
0.7048
SEEGNET
0.5919
0.7339
0.5744
0.6946
0.8076
0.6364
Appendix
Table 12: Ablation on resampling rate for pathological brain region identification (500 Hz).
Figure 4: Model interpretation analysis using SHAP.
Model
Precision
Recall
F1
Specificity
AUC
Outcome
SEEG-NET
0.5790
0.7169
0.5259
0.6049
0.7850
0.5952
CLAP
0.5936
0.6997
0.6009
0.7823
0.7684
0.6770
TimeConv-CNN
0.6259
0.7454
0.6469
0.8230
0.8061
0.7380
EEGNet (adapted)
0.5792
0.6957
0.5640
0.7075
0.7468
0.5942
Linear baseline
0.5323
0.5045
0.4899
0.9880
0.5045
0.4920
Appendix
Table 13: Additional baselines for pathological brain region identification (Task 2).
Patient Name
Dataset
Age
Gender
SO
SOZ
Res
Anat
Ev
Int
Ict
Awk
Split
sub-hupHUP060
hup
42
F
N
Y
Y
N
N
Y
Y
N
train
sub-hupHUP064
hup
21
M
Y
Y
Y
N
N
Y
Y
N
test
sub-hupHUP065
hup
36
M
Y
Y
Y
N
N
Y
Y
N
train
sub-hupHUP070
hup
33
M
Y
Y
Y
N
N
Y
Y
N
train
sub-hupHUP074
hup
25
F
Y
Y
Y
N
N
Y
Y
N
train
sub-hupHUP075
hup
57
F
N
Y
Y
N
N
Y
Y
N
train
Appendix
Table 14: Summary of patient metadata and recording availability. Column abbreviations: SO = Surgical Outcome; SOZ = Seizure Onset Zone Annotation; Res = Resection Annotation; Anat = Anatomical Annotation; Ev = Event Annotation; Int = Interictal Recording; Ict = Ictal Recording; Awk = Awake Recording. Value abbreviations: M = Male; F = Female; Y = Yes (present); N = No (absent); - = Not applicable or not available.
Electroencephalography (EEG) studies brain-generated electrical activity to understand cognition and support diagnosis, monitoring, and brain-computer interfaces.
Accurate identification of the epileptogenic zone (EZ) is essential for seizure freedom after resective surgery in drug-resistant epilepsy, yet seizure freedom rates remain below 50%. We developed EpiiSLM, a dual foundation model system for EZ identification with stereo-electroencephalography (sEEG), by training a signal foundation model on 104,990 minutes of sEEG recordings from the Montreal Neurological Institute & Hospital, while leveraging all recordings regardless of surgical outcome and anchoring EZ biomarker extraction on non-epileptic signals. A language foundation model then integrates sEEG-derived outputs with multimodal clinical information to produce interpretable predictions. Under leave-one-patient-out evaluation, EpiiSLM achieved 0.978 contact-level positive predictive value (PPV), outperforming the seizure onset zone(SOZ)-as-EZ baseline by 15.1% (p < 0.05), and 100% region-level accuracy; on an external dataset, EpiiSLM achieved 0.857 contact-level PPV. EpiiSLM requires only one night of interictal sleep data, suggesting potential to reduce invasive sEEG monitoring duration and improve surgical outcomes.
Thi Kieu Khanh Ho, Thomas Lai, Petr Klimes +5
Department of Electrical and Computer Engineering, McGill University, Montr´eal, Quebec, Canada. · Mila-Quebec AI Institute, Montr´eal, Quebec, Canada. · Montreal Neurological Institute and Hospital, McGill University, Montr´eal, Quebec, Canada. +5
Clinicians and neuroscientists have long analyzed intracranial electroencephalography (iEEG) through directly measurable physiological characteristics, which carry much of the information that downstream tasks depend on. Recent iEEG foundation models learn by reconstructing or predicting their inputs, which leaves the retention of these characteristics implicit. They are also evaluated mainly on cognitive decoding and a narrow clinical task, i.e., seizure detection. On a broad, clinically relevant benchmark such as Omni-iEEG, they remain below task-specific models when used frozen. We introduce PHASE, a physiology-guided foundation model that makes these characteristics explicit learning targets, pairing them with masked latent prediction in a temporal stage (PHASE-T) within each channel and a spatiotemporal stage (PHASE-ST) across synchronized channels. PHASE is pretrained on heterogeneous recordings from 222 participants at nine clinical sites. On all five Omni-iEEG clinical tasks, frozen PHASE-T outperforms every evaluated foundation model by up to 31%, and fine-tuned PHASE-T surpasses the task-specific models, setting a new state of the art. PHASE-T benefits from physiological supervision, outperforming variants trained with latent prediction alone or auxiliary waveform reconstruction on every task in matched ablations. PHASE-T generalizes to unseen institutions, outperforming the compared models with few or no local labels. PHASE-ST further improves seizure-onset-zone identification over PHASE-T and, when frozen, decodes sound volume and pitch on BrainTreebank better than published models. Beyond task performance, PHASE learns to encapsulate the physiological characteristics clinicians recognize, from seizure onset and its propagation to anatomical region identity, even though its pretraining contains no ictal recordings or anatomical labels.
Yipeng Zhang, Chenda Duan, Yuanyi Ding +10
UCLA Samueli School of Engineering · UCLA Mattel Children’s Hospital, David Geffen School of Medicine · Children’s Hospital of Michigan, Wayne State University School of Medicine +1
Deep learning for EEG-based seizure detection faces critical challenges: severe annotation scarcity and extreme class imbalance, where ictal events comprise less than 10% of clinical recordings. We present DiffEEG, a 9.6M-parameter self-supervised foundation model that addresses both limitations through denoising diffusion pre-training and reinforcement learning (RL)-based fine-tuning. Pre-trained on 1.3M unlabeled segments from the Temple University Hospital Seizure Corpus (TUHSZ), DiffEEG learns generic neural representations via a 1D U-Net with multi-head self-attention. For downstream adaptation, a reinforced decision layer employs policy gradient optimization to directly maximize F1-score, prioritizing sensitivity to rare seizure events over overall accuracy. Under strict patient-wise evaluation (279 patients, Leave-One-Fold-Out), DiffEEG achieves 61% accuracy and 59% F1 for 4-class seizure subtyping, and 81% accuracy with 85% weighted F1 for binary detection, maintaining clinically viable seizure recall (59%) despite extreme imbalance (6.7% prevalence). Segment-level evaluation establishes an upper bound of 97.6% accuracy, confirming strong architectural capacity. DiffEEG demonstrates that diffusion-based pre-training combined with metric-aware reinforcement learning enables clinically deployable seizure monitoring with minimal labeled data requirements.
Abdulkader Helwan, Lina Abou-Abbas, Hussein El Amouri +2
Department of Electrical and Computer Engineering, Lebanese American University, Byblos, Lebanon · Institute of Applied Artificial Intelligence, TÉLUQ University, Montreal, Canada