Bayesian Uncertainty-aware Deep Learning with noisy labels: Tackling annotation ambiguity in EEG seizure detection

Clinical Snapshot

Key Findings

• P Value: Not reported in abstract

• Effect Size: Not reported in abstract; specific performance metric improvements not quantified

• Primary Outcome: BUNDL consistently identifies noisy labels and improves robustness of three base deep learning models under various label noise conditions across TUH, CHB-MIT, and Siena EEG datasets

• Nnt Or Sensitivity: Sensitivity, specificity, and F1-score values not provided in abstract; seizure onset zone localisation accuracy improvement reported qualitatively only

• Confidence Interval: Not reported in abstract

Clinical Application

The model-agnostic, parameter-free design of BUNDL is a significant practical advantage — it can theoretically be integrated into existing deep learning EEG pipelines without architectural modification. However, clinical deployment requires regulatory approval, prospective validation on local datasets, integration with existing EEG review software (e.g., Persyst, Natus), and clinician training. Computational cost analysis in the paper supports feasibility, but real-world implementation complexity is not addressed. In Australia, automated EEG seizure detection is used as a clinical adjunct in tertiary epilepsy centres and paediatric neurology units. The TGA would require prospective clinical validation data before regulatory clearance of any AI-based EEG diagnostic tool as a Class IIb or Class III medical device under the TGA Software as a Medical Device (SaMD) framework. The RACGP and Epilepsy Society of Australia do not currently have specific guidelines for AI-assisted EEG interpretation. PBS funding for epilepsy management (including anti-seizure medications and presurgical evaluation) would be indirectly affected by improvements in diagnostic accuracy. Australian EMUs, including those at Austin Health, Royal Melbourne Hospital, and Sydney Children's Hospital Network, use predominantly US and European EEG systems, suggesting technical compatibility, but local dataset validation remains essential before clinical adoption. Patients undergoing scalp EEG monitoring for epilepsy evaluation, including those in epilepsy monitoring units (EMUs), paediatric epilepsy services, and presurgical evaluation programmes. Most directly applicable to institutions using automated EEG review systems as adjuncts to clinical interpretation.

Abstract

Deep learning is advancing EEG processing for automated epileptic seizure detection and onset zone localization, yet its performance relies heavily on high-quality annotated training data. However, scalp EEG is susceptible to high noise levels, which in turn leads to imprecise annotations of the seizure timing and characteristics. This "label noise" presents a significant challenge in model training and generalization. In this paper, we introduce Bayesian UncertaiNty-aware Deep Learning (BUNDL), a novel algorithm that informs a deep learning model of label ambiguities, thereby enhancing the robustness of seizure detection systems. By integrating domain knowledge into an underlying Bayesian framework, we derive a novel KL-divergence-based loss function that capitalizes on uncertainty to better learn seizure characteristics from scalp EEG. Thus, BUNDL offers a straightforward and model-agnostic method for training deep neural networks with noisy training labels that does not add any parameters to existing architectures. Additionally, we explore the impact of improved detection system on the task of automated onset zone localization. We validate BUNDL using a comprehensive simulated EEG dataset and two publicly available datasets collected by Temple University Hospital (TUH) and Boston Children's Hospital (CHB-MIT). Results show that BUNDL consistently identifies noisy labels and improves the robustness of three base models under various label noise conditions. We also conduct ablation experiments on uncertainty quantification, evaluate cross-site generalizability to Siena EEG dataset, and quantify computational cost of all methods. Furthermore, we demonstrate that BUNDL improves seizure onset zone localization accuracy. Ultimately, BUNDL presents as a reliable method that can be seamlessly integrated with existing deep models used in clinical practice, enabling the training of trustworthy models for epilepsy evaluation.

References

1. 1.Shama, M., Deeksha, & Venkataraman, A. (2026). Bayesian Uncertainty-aware Deep Learning with noisy labels: Tackling annotation ambiguity in EEG seizure detection. PLOS ONE. https://doi.org/10.1371/journal.pone.0352191