📊 statistics

Bayesian uncertainty-aware deep learning with noisy labels: Tackling annotation ambiguity in EEG seizure detection

This paper introduces BUNDL, a novel, model-agnostic Bayesian deep learning algorithm that utilizes a KL-divergence-based loss function to handle label noise in EEG seizure detection, thereby improving model robustness and generalizability without adding extra parameters.

Original authors: Deeksha M. Shama, Archana Venkataraman

Published 2026-06-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Deeksha M. Shama, Archana Venkataraman

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Core Problem: Teaching with a "Bad Textbook"

Imagine you are a teacher trying to train a student (an AI computer program) to recognize epileptic seizures from brain wave recordings (EEG). To do this, you give the student a textbook filled with examples of seizures and normal brain activity.

However, there is a major problem: The textbook is written by humans who make mistakes.

In the medical world, doctors look at these complex brain waves and mark where a seizure starts and ends. But because the signals are noisy and hard to read, doctors often disagree. Sometimes they mark a seizure as starting earlier than it actually did (over-segmentation), or they miss the exact end time. This is called "label noise."

If you teach your AI student using this imperfect textbook, the student will learn the mistakes. It will think that "noise" is part of a seizure, or it will get confused about when a seizure actually happens. This makes the AI unreliable.

The Solution: BUNDL (The "Skeptical Student")

The authors created a new training method called BUNDL (Bayesian Uncertainty-aware Deep Learning).

Think of BUNDL not as a new type of student, but as a new teaching strategy that makes the student smarter about when to trust the textbook.

Here is how it works, using an analogy:

The "Confidence Check" (Monte Carlo Dropout):
Normally, when the AI looks at a brain wave, it gives one answer: "This is a seizure."
With BUNDL, the AI is forced to look at the same brain wave multiple times, but each time it ignores a random part of its own brain (a technique called dropout).
- If the AI says "Seizure" every single time, it is confident.
- If the AI says "Seizure" sometimes and "Normal" other times, it is uncertain.
The "Trust Meter" (Uncertainty as a Proxy for Noise):
The authors realized that when the AI is uncertain about a specific brain wave, it’s usually because the human label in the textbook is wrong or ambiguous.
- High Confidence: The AI trusts the textbook label.
- High Uncertainty: The AI thinks, "Wait, I’m confused. The textbook says this is a seizure, but my own analysis is shaky. The textbook might be wrong here."
The "Smart Correction" (The Loss Function):
Instead of blindly accepting the textbook’s answer, BUNDL uses a mathematical formula to adjust the lesson.
- If the AI is uncertain, it leans less on the human’s label and more on its own pattern recognition.
- It essentially says: "I will learn from this example, but I won’t trust the human’s timestamp 100% because the data looks messy."

Why Is This Better Than Other Methods?

There are other ways to handle bad textbooks, but they have downsides:

The "Editor" Approach (Pruning): Some methods try to throw away the "bad" pages of the textbook before teaching. But in EEG data, it’s hard to know which pages are truly bad, so you might throw away important information.
The "Translator" Approach (Noise Layers): Some methods add a complex extra layer to the AI to "translate" the noisy labels into clean ones. This makes the AI much bigger, slower, and harder to build.

BUNDL is different because:

It’s Lightweight: It doesn’t add any new parts to the AI. It just changes how the AI learns from the existing data.
It’s Universal: It can be plugged into almost any existing AI model (like a universal adapter).
It’s Efficient: It doesn’t require storing huge amounts of past data or running multiple separate training stages.

The Results: Did It Work?

The researchers tested BUNDL on three things:

Fake Data: They created computer-simulated brain waves where they knew exactly where the errors were.
Real Hospital Data (TUH & CHB-MIT): They used real patient records from two different hospitals.
Unseen Data (Siena): They tested if the AI could work on data from a third hospital it had never seen before.

The Findings:

Fewer False Alarms: The biggest win was that BUNDL significantly reduced False Positives. In medical terms, this means the AI stopped crying "Wolf!" when there was no seizure. This is crucial because doctors hate being woken up by false alarms.
Better Localization: The AI got better at pinpointing where in the brain the seizure started (Seizure Onset Zone). This is important for surgeons who need to know exactly which part of the brain to operate on.
Robustness: Even when the human labels were very messy (over-segmented), BUNDL kept the AI performing well, whereas other methods struggled.

The Caveats (What the Paper Admits)

It’s a Training Tool, Not a New Brain: BUNDL isn’t a new AI model; it’s a way to train existing models better.
Sensitivity Trade-off: In some cases, by being so careful about false alarms, the AI missed a few actual seizures (lower sensitivity). The authors note this is likely because the human labels themselves were imprecise, so the AI was right to be skeptical.
Computational Cost: Because the AI has to look at the data multiple times to check its confidence, training takes a bit longer than standard methods. However, once trained, the AI runs at normal speed.

Summary

Imagine trying to learn to drive from a GPS that sometimes gives wrong directions.

Old Way: You trust the GPS blindly and crash.
Other New Ways: You hire a second GPS to argue with the first one (expensive and slow).
BUNDL Way: You learn to notice when the GPS is hesitating or giving vague directions. When it hesitates, you rely more on your own eyes and street signs. You don’t need a second GPS; you just need to be smarter about when to trust the first one.

BUNDL makes AI "skeptical" of bad human labels, leading to more reliable seizure detection without needing to rebuild the AI from scratch.

Technical Summary: Bayesian Uncertainty-aware Deep Learning with Noisy Labels for EEG Seizure Detection

Problem Statement
Deep learning models for automated epileptic seizure detection and onset zone localization rely heavily on high-quality annotated training data. However, scalp EEG signals are complex and artifact-prone, leading to imprecise clinical annotations regarding seizure timing and characteristics. This "label noise" poses a significant challenge, as inter-rater agreement among clinicians is often low (approximately 60%), and clinicians frequently over-segment seizures to minimize the risk of missing events. Standard supervised learning paradigms, which minimize cross-entropy loss against these noisy "ground truth" labels, cause models to learn misleading features, resulting in poor generalizability and performance plateaus. Existing methods for handling label noise often require architectural changes that add significant parameters, rely on unavailable prior information, or assume linear data variations that do not hold for complex EEG data.

Methodology: BUNDL Framework
The authors introduce Bayesian UncertaiNty-aware Deep Learning (BUNDL), a model-agnostic training strategy that integrates domain knowledge into a Bayesian framework to handle label ambiguity without adding parameters to existing neural network architectures.

Probabilistic Formulation: BUNDL distinguishes between observed noisy labels ( $y_n$ ) and unobserved clean labels ( $y_c$ ). It models the clean label posterior distribution $p(y_c|x)$ as a Bernoulli distribution that balances the clinician-provided label ( $p_{y_n}$ ) and the model’s own prediction ( $f_\theta(x)$ ), weighted by an unreliability measure $z_{x y_n}$ .
Uncertainty Quantification: The core innovation is the estimation of $z_{x y_n}$ , which quantifies the unreliability of the input data and given label. The authors hypothesize that label noise increases the entropy of the model’s output due to the "memorization effect" (where models fit clean samples first and remain uncertain about noisy ones). BUNDL uses Monte Carlo (MC) Dropout during training to estimate this uncertainty. By performing multiple forward passes ( $N=20$ ), it computes the average entropy of the predictions. High entropy (high $z_{x y_n}$ ) indicates the model should trust its own predictions over the noisy label; low entropy indicates the label is likely reliable.
Loss Function: The framework derives a novel KL-divergence-based loss function. Instead of minimizing the distance between predictions and noisy labels, it minimizes the KL divergence between the network’s prediction distribution and the estimated clean label posterior distribution. This allows the model to adaptively correct for noisy annotations during training.
Training Strategy: The method involves a pretraining phase on samples with high confidence (clearly seizure or baseline) to establish baseline uncertainty estimates, followed by fine-tuning using the BUNDL loss. The approach is fully automatic and requires only the clinician-provided labels.

Key Contributions

Uncertainty-Driven Label Correction: A probabilistic training framework that models uncertainty-informed label transitions to handle multiple forms of label noise (symmetric and asymmetric) without architectural modifications.
Model-Agnostic Algorithms: Algorithms that can be integrated into any deep neural network (tested on DeepSOZ, TGCN, and CNN architectures) with convergence analyses showing robustness across different noise settings.
Simulated EEG Benchmarks: The use of the SEREEGA framework to create a standardized pipeline for generating simulated EEG data with controlled, known label noise (symmetric, over-segmentation, and under-segmentation), enabling rigorous evaluation against ground truth.
Comprehensive Evaluation: Validation across simulated data and three real-world datasets (TUH, CHB-MIT, and Siena), including an analysis of cross-site generalizability and the impact on seizure onset zone (SOZ) localization.

Results

Simulated Data: BUNDL consistently outperformed baseline methods (Self-Adaptive Learning, Noisy Adaptation Layer, and standard Cross-Entropy Loss) across all noise conditions (symmetric, 10%/30% over-segmentation, and 10%/30% under-segmentation). It achieved the best Area Under the Receiver Operating Characteristic Curve (AUROC) and Area Under the Precision-Recall Curve (AUPRC) while significantly reducing false positive rates (FPR). Ablation studies confirmed that MC Dropout-based uncertainty was more effective than ensemble methods, test-time augmentation, or loss-based metrics for identifying label mismatch.
Real-World Data (TUH and CHB-MIT): BUNDL achieved the lowest false positive rates across all tested deep networks. On the TUH dataset, BUNDL with DeepSOZ achieved an FPR of 2.3 min/hour, significantly lower than baselines. On the CHB-MIT dataset, BUNDL with CNN achieved an FPR of 0.8 min/hour. While sensitivity varied, BUNDL consistently improved latency and robustness compared to methods that did not account for label noise.
Cross-Site Generalization: On the unseen Siena dataset, BUNDL achieved the lowest FPR among all methods, demonstrating robustness to distribution shifts, although sensitivity dropped, reflecting the challenges of out-of-distribution testing and potential label over-segmentation in the target dataset.
SOZ Localization: Incorporating BUNDL for seizure detection improved the accuracy of seizure onset zone localization in the TUH dataset (accuracy increased from 0.591 to 0.620 at the patient level), suggesting that cleaner detection labels lead to better localization performance.
Computational Cost: BUNDL requires no additional parameters. While it incurs a computational cost due to multiple forward passes for MC Dropout, it is more efficient than multi-stage training methods like NAL and does not require storing past predictions like Self-Adaptive Learning.

Significance
The paper claims that BUNDL presents a reliable, resource-efficient, and model-agnostic method for training deep neural networks with noisy labels. By leveraging Bayesian uncertainty quantification via MC Dropout, it allows models to implicitly identify unreliable annotations and adjust their learning process, thereby enhancing the robustness and generalizability of automated seizure detection systems. The authors emphasize that this approach can be seamlessly integrated into existing clinical deep learning pipelines, enabling the training of more trustworthy models for epilepsy evaluation without the need for additional manual annotation effort or complex architectural changes.