English abstract
Epilepsy is a neurological disorder that affects millions of individuals worldwide. The
recurrent seizures associated with epilepsy significantly affect the quality of life of
these patients. Accurate and timely seizure prediction will improve epilepsy
management and enhance these patients' quality of life. In this thesis, a novel non patient-specific seizure prediction system is proposed using the graded spiking neural
networks (GSNNs) implementable on Intel's Loihi 2 neuromorphic processor.
Trained using the widely used CHB-MIT EEG dataset, the GSNN-based EEG seizure
predictor achieved a very competitive performance, compared to existing EEG seizure
prediction methods. To further enhance the quality of the proposed seizure predictor,
The following tasks are performed: 1) optimized the predictor's hyperparameters, 2)
reduced the amount of processed EEG data by selecting the optimal subset of EEG
channels, and 3) incorporated a time-windowed voting mechanism to enhance the
system's robustness to noise and artifacts.
Finally the seizure prediction system is implemented on Intel's Loihi 2 neuromorphic
processor and analyzed its performance. A comparison between the proposed system
and those based on traditional artificial neural networks (ANNs) highlighted the
potential of GSNN-based system for greater efficiency and lower computational
complexity.
