Abstract
Epilepsy is one of the most common neurological disorders. Seizure prediction for patients with refractory epilepsy can alert patients to interventions and prevent many serious consequences. Aiming at the problem that most of the current epilepsy prediction algorithms are not suitable for hardware implementation into low-latency and low-power wearable or portable medical devices because of their high complexity and large number of parameters, this paper proposes a lightweight and hardware-friendly deep learning network, 1D-CNN-GRU model. The raw EEG data can be fed into the network for automatic feature extraction and classification after simple filtering and normalization. After fixed-point quantization and compression, the overall size of the model is only 6.955 KB. The proposed method has been evaluated on 23 samples from the scalp-EEG based CHB-MIT dataset provided by the Boston Children's Hospital-MIT. Experimental results demonstrate that the proposed model can achieve an average sensitivity of 94.63% and accuracy of 93.45% in the binary classification task of the pre-seizure 30 min signal and inter-seizure signal, and its lightweight feature fulfills the requirements for hardware implementation as a low-power, wearable epilepsy prediction medical device.