Channel-annotated deep learning for enhanced interpretability in EEG-based seizure detection

Currently, electroencephalogram (EEG) provides critical data to support the diagnosis of epilepsy through the identification of seizure events. The review process is undertaken by clinicians or EEG specialists and is labour-intensive, especially for long-term EEG recordings. Deep learning (DL) has been proposed to automate and expedite the seizure review and annotation process, providing superior performance when compared to traditional machine learning (ML) methods. However, DL algorithms lack interpretability which is a crucial factor for clinical adoption. Consequently, the "black-box"nature of these DL algorithms limits the transparency of these algorithms, preventing clinicians from having knowledge of how the predictions are derived. In this study, we propose a novel two-block seizure detection algorithm that leverages the channel annotation of the EEG recordings in the TUH EEG Seizure Corpus (TUSZ) based on the likelihood of seizure activities on each channel. This method allows direct interpretation of the EEG segment without requiring any further interpretability or Explainable Artificial Intelligence (XAI) methods during the prediction phase. Further, we adopted an explainable method for explaining decisions made by the seizure detection algorithms, identifying channels that influence the final predictions. This novel DL approach utilizing CNN, transformer and MLP achieved an AUC of 0.93, accuracy of 0.88, specificity of 0.88 and sensitivity of 0.82 for the seizure detection task, comparable with other state-of-the-art algorithms. Our algorithm was further validated on a separate continuous EEG dataset achieving an AUC of 0.82, accuracy of 0.72, specificity of 0.72 and sensitivity of 0.82. Additionally, we also evaluated the reliability and efficacy of our XAI method on predicted seizure events, achieving a sensitivity of 0.59 inaccurately localizing channels with seizure activities.