A 13.8mW ANN-Based Seizure Prediction Accelerator

Seizure occurrences can be very difficult to detect at a young age and may increase in severity with age. Currently, there is no patient-friendly seizure prediction solution that can precisely detect a seizure's occurrences to help prevent the injury of a patient. This paper presents a Neural Network (NN)-based classifier design to detect and predict seizures from EEG signals. The proposed system is divided into two phases, 1) training of the Classifier to detect/predict the seizure events, and 2) realization of the trained NN model for real-time implementation. The implemented exponent-mantissa sigmoid activation function in the Artificial NN (ANN) hidden layer exploits the IEEE-754 property and rearranges to attain an energy consumption of 0.39 µ J/operation and decreases the area by 28.2% compared to the conventional activation function realizations. The proposed system predicts the seizure with an average accuracy of 92.47%, 92.31%, and 88.77%, for 10 sec, 20 sec, and 30 sec, respectively, using the CHB-MIT EEG database with a latency of 1s.