Epileptic Seizure Prediction Using Spiking Neural Networks
Abstract number :
1.479
Submission category :
2. Translational Research / 2B. Devices, Technologies, Stem Cells
Year :
2024
Submission ID :
1436
Source :
www.aesnet.org
Presentation date :
12/7/2024 12:00:00 AM
Published date :
Authors :
Presenting Author: Gladys Addai-Domfe, BS – Augusta University
Hisham Daoud, PhD – Augusta University
Rationale: Epileptic seizure prediction remains a critical challenge in epilepsy research due to the unpredictable nature of seizures. While artificial intelligence (AI) has shown promising results in seizure prediction, current AI methods that are based on machine learning (ML) or deep learning (DL) face significant obstacles in terms of energy efficiency, computational costs, or accuracy. These limitations pose hurdles in developing an AI-powered wearable device for seizure prediction, which must operate within the constraints of embedded systems with limited computational resources. To address these challenges, this research proposes leveraging Spiking Neural Networks (SNNs), brain-inspired artificial neural networks that mimic the behavior of neurons in the brain, in developing a seizure prediction model as a step towards an energy-efficient and high-accuracy seizure prediction wearable device.
Methods: We formulated our problem as a classification task between preictal (1 hour before seizures) and interictal (at least 4 hours before/after seizures) brain states in which the seizure is predicted upon detecting the preictal state among the prevalent interictal state. In our experiment, we used the EEG dataset from CHB-MIT. Eight patients who met our timing criteria were selected for our experiment. EEG recordings were pre-processed and segmented to make up labelled preictal and interictal segments for model training. Spike encoding of the EEG segments is achieved using hard threshold applied on the leaky integrate and fire neuron model. The generated spike trains are then fed to our model which comprises a 3-layer Deep Spiking Convolutional Neural Network followed by a fully connected layer in which Sigmoid function is applied for predictions. Surrogate gradient descent is employed as our supervised learning algorithm which propagates error gradients back through the network and the network’s parameters are adjusted accordingly to improve the model prediction.
Results: To assess the performance of our model, we utilized the leave-one-out cross-validation testing strategy and calculated the prediction accuracy. Our approach demonstrated noteworthy performance, achieving an average classification accuracy of 87%. This outcome is in line with, or even exceeds, the performance achieved by traditional ML approaches (accuracy < 90%). Although the accuracy of our model is lower than the state-of-the-art DL methods (accuracy > 90%), its efficiency in computation and communication surpasses those that are based on DL. Thus, our method facilitates the realization of a wearable device that can consume low power while predicting seizures in real-time.
Conclusions: By employing SNNs, our research addresses key limitations of current ML/DL methods, including energy efficiency and handling temporal dynamics of EEG signals in predicting epileptic seizures. Our method holds promise for accurate seizure prediction in real-time, with potential integration into wearable devices for enhanced epileptic patients’ safety and improved quality of life.
Funding: AES BRIDGE Summer Research Internship Grant.
Translational Research