Epileptic Seizure Detection in Neonatal EEG Using a Multi-Band Graph Neural Network Model

Neonatal seizures are the most common clinical presentation of neurological dysfunction, requiring immediate attention and treatment. Manual detection of seizure events from continuous electroencephalogram (EEG) recordings is laborious and time-consuming. In this study, a novel graph-based method for automated neonatal seizure detection is proposed. The proposed method aims to improve the detection performance by the thorough representation of multi-channel EEG signals and the adaptive classification of multi-band graph representations. To achieve this, a band-wise feature extraction method is performed on the raw EEG to provide more detailed information for classification. In addition, a novel classification model, namely the multi-band graph neural network (MBGNN), is proposed, which utilizes the attention mechanism and can take full advantage of the multi-band graph representations to improve the classification performance. The proposed method is evaluated using the EEG recordings of 39 neonates from the Helsinki database. The MBGNN model gives an average area under the receiver operating characteristic curve (AUC) of 99.11%, an average positive predictive value (PPV) of 95.34%, and an average negative predictive value (NPV) of 96.66%. The experimental results show that the proposed method could fully exploit the multi-band EEG information and facilitate the classification of seizure/non-seizure EEG epochs, making it more appealing for patient-specific clinical applications.