Explainable classification of seizures and other patterns of harmful brain activity in critically ill patients

Abstract

Accurate detection and classification of seizures from electroencephalography (EEG) signals play a vital role in enabling timely medical interventions and treatments for neurological disorders. Currently, EEG recordings are analyzed exclusively by trained human experts. Although essential, this manual process is time-intensive, costly, and prone to fatigue-related errors, in addition to inconsistencies between reviewers. In this work, we propose a deep neural network (DNN) model equipped with interpretable layers designed to classify seizures and other abnormal brain patterns, such as periodic discharges, rhythmic delta activity, and miscellaneous pathological events. The proposed DNN architecture incorporates explainable components that allow clinicians to trace and understand the model’s reasoning process, fostering confidence and aiding clinical decision-making. This combination of deep learning with interpretability mechanisms is novel and overcomes several limitations of traditional approaches. The method is validated using a publicly available EEG dataset, achieving state-of-the-art accuracy while maintaining interpretability that supports expert review. This study contributes to the growing field of EEG-based seizure detection by bridging the gap between high-performance DNN-based models and the need for clinical transparency. Unlike existing methods, our approach employs an ensemble of three specialized DNN, each capturing complementary feature representations. This ensemble framework improves both interpretability and robustness of the classification results.