Universal EEG Epilepsy Detection via Evidential Multi-View De-Biasing

Ziqi Wen, Cai Xu, Wanqing Zhao, Jie Zhao, Wei ZhaoAAAI 2026 · DOI

Abstract

Epilepsy is a widespread neurological disorder characterized by highly patient-specific EEG patterns. Existing EEG-based seizure detection methods either train individualized models for each patient or adapt models pre-trained on known patients to new ones. However, when encountering previously unseen patients, these methods typically require retraining or fine-tuning, which limits their practical utility in clinical settings. This limitation can be linked to biases caused by patient-specific variations, which obscure the underlying pathological patterns of seizures. To address this, we propose an evidential multi-view framework that reinforces the learning of core epileptic features by promoting consistency across multiple views and reducing reliance on high-uncertainty, patient-specific segments. Specifically, we introduce Bias-guided Fisher-Evidential Multi-View Learning (BF-EML) to guide the model toward discovering intrinsic seizure patterns. BF-EML employs a two-stage training architecture: In Stage 1, we use the Fisher Information Matrix to reorder EEG segments by uncertainty and deliberately train a biased feature generator on low-evidence segments. In Stage 2, we design a dual-branch network where the biased and unbiased branches are alternately trained, encouraging the unbiased branch to reduce its reliance on patient-specific biases. Finally, we introduce a shift-calibrated fusion strategy to enhance the consistency of pathogenic feature integration. Extensive experiments on public datasets and a clinical dataset demonstrate that our method achieves superior performance in both single- and multi-patient scenarios. Importantly, it generalizes well to unseen patients without the need for retraining.

Introduction

EEG-based seizure detection is critical for epilepsy management, yet it faces a significant hurdle: inter-patient variability. Seizure patterns are highly specific to individuals, often leading models to learn patient-specific "noise" rather than universal "pathology." Existing multi-patient methods typically rely on domain adaptation, which requires data from a new patient to "fine-tune" the model—a process that is often impractical in urgent clinical settings.

Comparison between existing multi-patient methods and the proposed BF-EML concept.

The proposed Bias-guided Fisher-Evidential Multi-View Learning (BF-EML) framework addresses this by identifying and suppressing patient-specific biases, allowing for a "universal" model that generalizes to unseen patients without retraining.

Methods

The BF-EML framework operates through a sophisticated two-stage training process designed to decouple universal seizure signatures from individual brain rhythm quirks.

The two-stage architecture of the BF-EML framework, featuring Fisher-Evidential re-ranking and dual-branch debiasing.

Stage 1: Fisher-Evidential Re-ranking

Standard evidential learning often fails to distinguish between varying levels of uncertainty. By incorporating the Fisher Information Matrix (FIM), the authors precisely quantify uncertainty. They found that "low-evidence" segments are typically the most biased by patient-specific artifacts. These segments are flagged to guide the debiasing process.

Stage 2: Dual-Branch Alternating Debiasing

The architecture employs two branches:

  1. Biased Branch: Intentionally trained on low-evidence samples to capture patient-specific quirks.
  2. Unbiased Branch: Trained on all data but forced to be statistically independent of the biased branch using the Hilbert-Schmidt Independence Criterion (HSIC). By explicitly modeling what the bias looks like, the system can effectively "subtract" it from the primary detection logic.

Shift-Calibrated Fusion

To integrate multiple EEG views (e.g., temporal and spectral), a shift-calibration mechanism ensures that the fusion process accounts for how debiasing affects different views differently, preventing high-confidence but potentially biased views from dominating the final decision.

Results

The framework was validated on the CHB-MIT benchmark, the Siena dataset, and a large-scale clinical dataset of 523 patients.

  • Superior Generalization: On the CHB-MIT dataset, BF-EML achieved 97.65% accuracy on unseen patients, significantly outperforming existing multi-patient methods.
  • Zero-Shot Clinical Utility: The model requires no retraining or fine-tuning for new patients, making it a "plug-and-play" solution for clinical environments.
  • Scalability: Performance improved as the training pool grew, demonstrating that the model effectively leverages patient diversity to refine its understanding of universal pathological features.

Conclusion

BF-EML represents a significant step toward universal epilepsy detection. By treating uncertainty as a signal for bias and using a dual-branch architecture to isolate it, the framework overcomes the patient-specific barrier. Future work will focus on optimizing the computational overhead of the Fisher Information Matrix and extending the model to real-time, low-latency applications for implantable seizure detection devices.

Summary updated

Abstract

Epilepsy is a widespread neurological disorder characterized by highly patient-specific EEG patterns. Existing EEG-based seizure detection methods either train individualized models for each patient or adapt models pre-trained on known patients to new ones. However, when encountering previously unseen patients, these methods typically require retraining or fine-tuning, which limits their practical utility in clinical settings. This limitation can be linked to biases caused by patient-specific variations, which obscure the underlying pathological patterns of seizures. To address this, we propose an evidential multi-view framework that reinforces the learning of core epileptic features by promoting consistency across multiple views and reducing reliance on high-uncertainty, patient-specific segments. Specifically, we introduce Bias-guided Fisher-Evidential Multi-View Learning (BF-EML) to guide the model toward discovering intrinsic seizure patterns. BF-EML employs a two-stage training architecture: In Stage 1, we use the Fisher Information Matrix to reorder EEG segments by uncertainty and deliberately train a biased feature generator on low-evidence segments. In Stage 2, we design a dual-branch network where the biased and unbiased branches are alternately trained, encouraging the unbiased branch to reduce its reliance on patient-specific biases. Finally, we introduce a shift-calibrated fusion strategy to enhance the consistency of pathogenic feature integration. Extensive experiments on public datasets and a clinical dataset demonstrate that our method achieves superior performance in both single- and multi-patient scenarios. Importantly, it generalizes well to unseen patients without the need for retraining.
BibTeX
@inproceedings{wen2026universal,
  author = {Ziqi Wen and Cai Xu and Wanqing Zhao and Jie Zhao and Wei Zhao},
  title = {{Universal EEG Epilepsy Detection via Evidential Multi-View De-Biasing}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2026},
  doi = {10.1609/aaai.v40i32.39896},
}
Figure