If artificial intelligence (AI) is to make accurate predictions, it needs to be trained on reliable information. But in medicine, even measurements and scans produced by clinicians contain uncertainties, making such data “fuzzy.”
Deeksha M. Shama, a PhD student in the Department of Electrical and Computer Engineering at Johns Hopkins University, along with Archana Venkataraman, associate professor at Boston University, have developed a new approach designed to address the problem.
Called Bayesian Uncertainty-aware Deep Learning (BUNDL), the method helps deep-learning models account for uncertainty in the labels of data used during training. In the case of people with epilepsy, better data labeling could make automated seizure detection more robust. The team’s paper was recently published in PLOS ONE.
“AI can help make seizure detection faster and more efficient,” says Shama. “But its performance can drop when the training data is not accurate. We wanted to address this issue by building AI systems that stay reliable even when the data is imperfect.”
That challenge is particularly significant in epilepsy care, where clinicians rely on electroencephalography (EEG) to monitor brain activity and identify seizures. EEG recordings can be complex and susceptible to noise, making it difficult to determine precisely when a seizure starts or stops. Clinicians may also interpret the same recording differently, creating uncertainty in the annotations used to train AI systems.
The researchers described these inaccuracies as “label noise.” In supervised deep-learning models, clinicians’ annotations are treated as what is known as “ground truth.” But EEG signals are notoriously difficult to interpret, and studies have found limited agreement among clinicians reviewing EEG data. Such disparities in clinicians’ evaluation of the timing and characteristics of seizures can therefore become part of the data used to train AI models.
Simulated EEG and label noise generation pipeline using SEREEGA. The top half shows components in source domain including source positions and five types of signals assigned to corresponding source indicating with following icons: background is shown in blue circles, the spike noise in purple triangle, and seizure source in red diamond. The background sources are fixed in position while the latter two is randomly assigned per simulated participant. The bottom half shows the scalp electrode plot of 10-20 montage and an example EEG with true seizure and noisy annotations marked.
When an AI system learns from those labels without accounting for their uncertainty, it can learn misleading patterns and have difficulty using the training data to evaluate new patients or different clinical settings.
“BUNDL takes a different approach,” Shama says. “Rather than treating every training label with equal certainty, the method uses uncertainty information to adjust how the model learns from the data itself.” The researchers designed BUNDL to work with existing deep-learning architectures without adding parameters to the underlying models, making the approach model-specific and more flexible.
The team also tested whether BUNDL could generalize to EEG data from a different source, using the Siena EEG dataset to evaluate how well the method performed across datasets.
Across the experiments, BUNDL improved the robustness of seizure-detection models under different types of label noise. On real-world datasets, the method also reduced false-positive rates by as much as 50% in some scenarios.
The researchers also examined whether BUNDL could improve downstream seizure-onset-zone localization—the automated identification of the region of the brain where a seizure begins. They found that BUNDL improved localization accuracy from 60.1% to 63.3% at the seizure level and from 59.1% to 62.0% at the patient level. “These results suggest that BUNDL could eventually provide useful information for epilepsy evaluation and treatment planning,” Shama says.
Although the team’s findings do not suggest the technology should replace clinical judgment, the approach points toward a broader goal: developing medical AI systems that are accurate, transparent, dependable, and better suited to the realities of clinical data. Because BUNDL is designed to work across different deep-learning architectures, the approach could potentially extend beyond epilepsy to other medical fields that rely on human annotations, including pathology and radiology.
“The next steps focus on making these AI systems easier to understand and trust,” Shama says. “Future work will include developing clearer, more human-readable explanations of how AI systems detect seizures and identify where they begin, aligning those explanations with clinical knowledge, and identifying and reducing potential sources of bias.”