Uncertainty-Aware Bayesian Deep Learning with Noisy Training Labels for Epileptic Seizure Detection.

Supervised learning has become the dominant paradigm in computer-aided diagnosis. Generally, these methods assume that the training labels represent "ground truth" information about the target phenomena. In actuality, the labels, often derived from human annotations, are noisy/unreliable. This aleoteric uncertainty poses significant challenges for modalities such as electroencephalography (EEG), in which "ground truth" is difficult to ascertain without invasive experiments. In this paper, we propose a novel Bayesian framework to mitigate the effects of aleoteric label uncertainty in the context of supervised deep learning. Our target application is EEG-based epileptic seizure detection. Our framework, called BUNDL, leverages domain knowledge to design a posterior distribution for the (unknown) "clean labels" that automatically adjusts based on the data uncertainty. Crucially, BUNDL can be wrapped around any existing detection model and trained using a novel KL divergence-based loss function. We validate BUNDL on both a simulated EEG dataset and the Temple University Hospital (TUH) corpus using three state-of-the-art deep networks. In all cases, BUNDL improves seizure detection performance over existing noise mitigation strategies.