Epileptic Spike Detection by Recurrent Neural Networks with Self-Attention Mechanism
Automated identification of epileptiform discharges for the diagnosis of epilepsy can mitigate the burden of the exhaustive manual search in electroencephalogram (EEG). Recent studies have indicated that a two-step method that consists of detection of candidate waveforms with signal processing and pattern matching followed by machine learning-based classification is effective. However, the overall performance depends on the detector of candidates. This paper thus considers a scenario without candidate waveforms, that is, we propose a recurrent neural network (RNN)-based self-attention model that can be fitted from the EEG segments generated without detecting spike candidates. In comparison with the state-of-the-art machine learning models which can be applied for EEG classification (LightGBM and EEGNet), the proposed model achieved higher performance (average accuracy: 90.2 %). This result strongly suggests that the self-attention mechanism is suitable to an automated identification of the epileptiform discharge in the EEG.