Real-Time Seizure Prediction Using Low-Power Spiking Neural Networks on Wearable Edge Devices

Abstract

Epilepsy is a condition that is present in 50 million individuals all over the world, but predicting seizures in real-time on wearable devices is a challenging task because of stringent power and latency requirements. The current paper introduces a wearable edge devices-based low-power seizure prediction system based on Spiking Neural Networks (SNNs). In contrast to traditional deep learning models, SNNs compute EEG signals as a result of events, and spikes, which consumes much less energy by a factor of ten. We suggest a three-layer SNN, leaky integrate-and-fire neurons, adaptive thresholding and sparse connectivity, trained using a surrogate gradient descent. The system based on CHB-MIT scalp EEG database and patient-wise splitting, has a sensitivity of 88.4, specificity of 93.2 and a false alarm rate of 0.08/hour, predicting seizures 5-30 minutes before they occur. The system is implemented on an ARM Cortex-M4 microcontroller, using 3.4 - 13 times less energy than a CNN-LSTM baseline to make an inference, and has 45 ms latency. Ablation experiments validate the work of adaptive thresholding, population coding and 16 ms timestep resolution. This is the first SNN-based seizure prediction system to the best of knowledge that has been tested on a publicly available EEG dataset and with wearable-edge hardware simulation. The framework shows that neuromorphic computing has the capabilities to address clinical and engineering needs of real-time neurological monitoring.