Objective: The electroencephalogram (EEG) tool has great potential for real-time monitoring of abnormal brain activities, such as preictal and ictal seizures. Developing an EEG-based detection system for patients with epilepsy is vital for clinical management and targeted therapy.
Methods: This paper proposes a single-channel seizure detection system using brain-rhythmic recurrence biomarkers (BRRM) and an optimized model (ONASNet). BRRM is a direct mapping of the recurrence morphology of brain rhythms in phase space; it reflects the nonlinear dynamics of original EEG signals. The architecture of ONASNet is determined through a modified neural network searching strategy. Then, we exploited transfer learning to apply ONASNet to our EEG data. The combination of BRRM and ONASNet leverages the multiple channels of a neural network to extract features from different brain rhythms simultaneously.
Results: We evaluated the efficiency of BRRM-ONASNet on the real EEG recordings derived from Bonn University. In the experiments, different transfer-learning models (TLMs) are respectively constructed using ONASNet and seven well-known neural network structures (VGG16/VGG19/ResNet50/InceptionV3/DenseNet121/Xception/NASNet). Moreover, we compared those TLMs by model size, computing complexity, learning capability, and prediction latency. ONASNet outperforms other structures by strong learning capability, high stability, small model size, short latency, and less requirement of computing resources. Comparing BRRM-ONASNet with other existing methods, our work performs better than others with 100% accuracy under the identical dataset and same detection task. Contributions: The proposed method in this study, analyzing nonlinear features from phase-space representations using a deep neural network, provides new insights for EEG decoding. The successful application of this method in epileptic-seizure detection contributes to computationally medical assistance for epilepsy.