Source-linked AI summary
EEGBind: Detecting Source-Level Interictal Epileptiform Discharges via EEG-Centric Multimodal Binding
Muchen Li, Anglin Liu, Xuetian Gao, Ruijian Xu, Jintai Chen
TL;DR
Source-level IED classification must assign IED-positive activity to clinically meaningful regions despite subtle, variable EEG evidence and imperfect multimodal context. EEGBind keeps EEG as the organizing modality, binds synchronized video as bounded auxiliary evidence, and applies view-consistent repair. It achieves 0.8395 weighted-F1 on the NMM-Source-IED benchmark and ranks first in the official evaluation.
Problem
Source-level IED analysis must distinguish five clinically meaningful source-region categories from subtle and variable evidence in short EEG windows.
Method
EEGBind uses EEG as the source anchor, binds synchronized video-context routes around it, and applies CE-free view-consistent repair under route perturbations.
Results
0.8395 weighted-F1 ranked the subject-disjoint five-fold EEGBind ensemble first in the official NMM-Source-IED evaluation.
Takeaways & Limitations
The results support EEG-centric multimodal binding as a practical strategy for source-level IED classification while keeping video in a bounded contextual role.
Takeaways & Limitations
The visualizations show sensor-space field structure rather than inverse source estimates.
Abstract
from arXiv · showhide
Source-level analysis of interictal epileptiform discharges (IEDs) is relevant to presurgical evaluation and treatment planning because it helps characterize where epileptiform activity is likely to arise. Beyond detecting whether an IED is present, this setting requires assigning IED-positive activity to clinically meaningful brain-region categories. This setting is challenging because source-region evidence in short electroencephalography (EEG) windows can be subtle, partial, and affected by subject variability, class imbalance, and imperfect multimodal context. We present EEGBind, an EEG-centric multimodal binding framework for five-class source-level IED classification. EEGBind treats EEG as the primary modality and binds synchronized video-context features around an EEG-centric representation. Instead of relying on early or overly strong multimodal fusion, which may perturb the source-sensitive EEG representation, EEGBind uses video context as auxiliary evidence for robust classification. A view-consistent repair stage is further used to improve hidden-set robustness while preserving the learned source-class boundary. On the NeuroMM 2026 Grand Challenge Track 3 NMM-Source-IED benchmark, EEGBind achieves 0.8395 on weighted-F1 and outperforms strong competitors. These results support EEG-centric multimodal binding as a practical strategy for source-level IED classification. The open-source code is available at https://github.com/HKUSTGZ-ML4Health-Lab/NeuroMM2026_IED_Detection.
1 Introduction
Source-level IED analysis extends detection by assigning epileptiform activity to clinically meaningful source regions, but subtle and variable EEG evidence makes this difficult. EEGBind addresses this with EEG-centric binding of synchronized video context and reports strong benchmark performance.
- Motivation: Source-level IED analysis supports localization relevant to presurgical evaluation, treatment planning, and interpretation of epileptiform patterns.
- Challenge: Short EEG windows contain subtle, partial, and distributed regional evidence that varies across subjects, conditions, and source regions.
- Approach: EEGBind uses EEG as the source anchor and synchronized video routes as bounded contextual evidence rather than an independent source classifier.
- Results: 0.8395 weighted-F1 ranked EEGBind first on the official hidden-test evaluation.The final subject-disjoint five-fold ensemble uses EEG-centric binding and CE-free repair.
2 Related Work
Prior work combines EEG with contextual or multimodal representations and develops methods for missing or perturbed modalities. EEGBind addresses the asymmetric case where EEG remains primary while auxiliary video context may be incomplete or unstable.
- EEG IED Analysis: Earlier IED analysis used morphology-sensitive signal processing and hand-designed descriptors, while EEG foundation models provide transferable representations for downstream tasks.
- Multimodal EEG: Prior multimodal EEG studies use video-derived motion and pose cues to add contextual evidence for clinical or binary IED detection.
- Robust Multimodal Learning: Robust multimodal methods address missing inputs through mask-aware attention, virtual representations, invariant features, masked pretraining, or resilient fine-tuning.
- Representation Consistency: Representation-consistency methods align perturbed views, but do not directly target a consistently available primary modality with incomplete auxiliary routes.
3 Methodology
EEGBind formulates source-level prediction as five-class classification with EEG providing the primary representation and synchronized video routes supplying auxiliary context. A CE-free repair stage stabilizes predictions and latent features under route perturbations, followed by subject-disjoint ensemble averaging.
- EEG-Centric Multimodal Binding: EEGBind predicts one of five source-region labels from official IED-positive windows using EEG as the primary source representation and video as auxiliary context.
- EEG-Centric Multimodal Binding: The binding module combines EEG features with synchronized video-context tokens to produce a candidate representation and source logits without an independent visual classifier.
- CE-Free View-Consistent Repair: The CE-free repair stage preserves the learned source-classification boundary while encouraging stable latent and logit behavior under missing or perturbed video routes.
- CE-Free View-Consistent Repair: The repair loss combines view-drop, feature-target, masked-route, and queue-based consistency terms.
- Subject-Disjoint Fold Ensemble: Subject-disjoint five-fold ensembling converts fold logits to log probabilities and averages them with equal weights.
4 Experiments
On the five-class NMM-Source-IED benchmark, EEGBind combines EEG-centric binding, auxiliary video context, repair, and subject-disjoint ensembling to improve source-level classification robustness.
- Official Evaluation Analysis: 0.8395 weighted-F1 was achieved by the final five-fold EEGBind ensemble, compared with 0.6444 for the EEG-only baseline.The repaired single model reached 0.8282, while Figure 2 visualizes the progression reported in Table 1.
- Official Evaluation Analysis: 0.7448 weighted-F1 was obtained with dummy video and 0.7460 when video was zeroed only at inference, below corresponding full-video configurations.These controls separate useful video-context binding from model-capacity effects.
- Implementation Details: EEGBind uses ST-EEGFormer-large, seven synchronized video routes, EEG-centric attentive binding, and one repair epoch without cross-entropy or binding losses.The final ensemble applies this pipeline across subject-disjoint folds and averages predictions.
- Implementation Details: EEGBind output features formed more separated source-class regions than overlapping ST-EEGFormer EEG anchor features on subject-disjoint out-of-fold projections.The representation map is complementary to the official hidden-test weighted-F1 evaluation.
- Implementation Details: 1.03% of predictions changed after zeroing all video routes, while single-route removal retained 99.68–100% agreement with the full view.The perturbation analysis covered 2,514 examples from 52 subjects and used subject-clustered bootstrap intervals.
- Repair Analysis: Repair variants increased weighted-F1 from 0.8131 to 0.8203, 0.8239, and 0.8282 as consistency, masked-route prediction, and contrastive terms were added.These results support the repair stage as a lightweight robustness step around the learned binding representation.
5 Discussion
EEGBind keeps EEG as the source anchor while using synchronized video as bounded context, with repair and controls targeting robustness without shifting the learned source boundary.
- 5 Discussion: EEGBind treats EEG as the primary modality and binds video-context routes around its source representation rather than using video as an independent source classifier.This design reflects the distinction between EEG-defined source labels and video-provided recording-condition or patient-state context.
- 5 Discussion: The CE-free repair stage improves consistency under route perturbations without introducing another supervised update to the learned source-class boundary.The repair strategy is applied to multimodal representations under route-dropped and masked video contexts.
- 5 Discussion: Dummy-video training isolates attentive-probe capacity, while zero-video inference measures dependence on trained video routes under the challenge protocol.These controls examine whether the video pathway contributes beyond model capacity alone.
- 5 Discussion: Official evaluation results support EEG-centric binding, lightweight CE-free repair, and subject-disjoint ensembling as the EEGBind pipeline.The discussion frames these components as the supported overall design while calling for stronger subject-heldout validation and bounded multimodal use.
6 Conclusion
EEGBind addresses source-level IED classification by binding video context around an EEG-centered representation and repairing route perturbations without disrupting source evidence. In official evaluation, the five-fold ensemble reaches 0.8395 weighted-F1, supporting the design as a practical strategy.
- 6 Conclusion: 0.8395 weighted-F1 is reached by the subject-disjoint five-fold ensemble, after multimodal binding improves the EEG-only baseline and repair further raises the single-model score.The reported progression is 0.6444 to 0.8019 with multimodal binding, then 0.8282 after repair, and 0.8395 for the ensemble.
- 6 Conclusion: EEGBind uses synchronized video as bounded candidate context around EEG source evidence, while CE-free view-consistent repair improves robustness under route perturbations.The source prediction remains anchored in the bound candidate representation rather than an independent visual classifier.
- 6 Conclusion: Dummy-video training and zero-video inference score 0.7448 and 0.7460, respectively, indicating that the gain is not explained by model capacity alone.The controls support meaningful video-context binding as part of the reported design.