Abstract
Accurate detection and classification of seizures from electroencephalography (EEG) data can potentially enable timely interventions and treatments for neurological diseases. Currently, EEG recordings are exclusively reviewed by human experts, namely neurologists with specialized training. While indispensable, this time-consuming workflow represents a major bottleneck. Review of EEG records is laborious, time-consuming, expensive, prone to fatigue-induced errors, and suffers from inter-rater reliability even among expert reviewers. This paper introduces a new deep neural network (DNN) with interpretable layers for the classification of seizures and other pathologic brain activities such as periodic discharges, rhythmic delta waves and miscellaneous activities. The DNN architecture uses interpretable layers that allow clinicians to evaluate the model’s decision-making pipeline and build trust in the model and support clinical decision making. The combination of deep learning and interpretability layers is novel and addresses the limitations of existing methods. We demonstrate the usefulness of the proposed approach on a publicly available EEG dataset. Our method achieves state-of-the-art performance and provides classification decisions that are interpretable, useful for clinical experts. This paper contributes to the existing body of literature on EEG-based seizure detection and addresses the gap between DNN-based methods and clinical interpretability, leading to accurate and clinically meaningful predictions.