Bayesian Uncertainty-aware Deep Learning with noisy labels: Tackling annotation ambiguity in EEG seizure detection
Deeksha M Shama1,2, Archana Venkataraman1,2
1Department of Electrical and Computer Engineering, Johns Hopkins University, Baltimore, Maryland, United States of America.
Plos One
|June 23, 2026
Summary
Bayesian UncertaiNty-aware Deep Learning (BUNDL) tackles noisy labels in electroencephalography (EEG) for epilepsy detection. This method enhances deep learning model robustness without adding parameters, improving seizure detection and localization accuracy.
Area of Science:
- Computational Neuroscience
- Medical Informatics
- Machine Learning
Background:
- Deep learning models for electroencephalography (EEG) analysis require high-quality annotated data for accurate epileptic seizure detection and onset zone localization.
- Scalp EEG data often suffer from high noise levels, leading to imprecise seizure annotations (label noise), which challenges model training and generalization.
Purpose of the Study:
- Introduce Bayesian UncertaiNty-aware Deep Learning (BUNDL), a novel algorithm designed to improve the robustness of deep learning models for seizure detection by accounting for label ambiguities.
- Develop a KL-divergence-based loss function within a Bayesian framework to leverage uncertainty for better learning of seizure characteristics from noisy EEG data.
- Evaluate the impact of BUNDL on automated seizure onset zone localization.
Main Methods:
- BUNDL integrates domain knowledge into a Bayesian framework, deriving a novel KL-divergence-based loss function to handle noisy training labels.
- The method is model-agnostic and does not introduce additional parameters to existing deep learning architectures.
- Validation was performed using simulated EEG data and two public datasets (TUH, CHB-MIT), with additional cross-site generalizability testing on the Siena EEG dataset.
Main Results:
- BUNDL effectively identifies noisy labels and enhances the robustness of three base deep learning models across various label noise conditions.
- Ablation studies confirmed the effectiveness of uncertainty quantification, and computational costs were evaluated.
- The algorithm demonstrated improved accuracy in seizure onset zone localization tasks.
Conclusions:
- BUNDL offers a straightforward and reliable method for training deep neural networks with noisy EEG data, enhancing trustworthiness.
- The algorithm can be seamlessly integrated into existing clinical deep learning models for epilepsy evaluation.
- Improved seizure detection robustness translates to better seizure onset zone localization accuracy.
