Towards Personalized Edge-AI for Medicine: Efficient Neuromorphic Frameworks for Seizure Detection and Prediction

Epilepsy affects approximately 1% of global population, with 30-40% of cases resistant to conventional pharmacological treatments. Neurostimulation offers these patients an alternative by delivering targeted electrical stimulation to the brain. However, adaptive AI for these devices relies on external or cloud-based resources for model retraining and updating, which introduces significant challenges, including high false positive rates, latency issues, and privacy concerns. We present a neuromorphic framework for real-time seizure detection and prediction, directly implemented on a neuromorphic SOC with on-chip learning capabilities. By leveraging spiking neural networks and few-shot edge learning, our system enables continual, patient-specific adaptation without requiring data transmission or cloud connectivity. For a detection framework, a model pre-trained on the TUH dataset is deployed on the BrainChip Akida processor and personalized using long-term EPILEPSIAE recordings. For prediction, we demonstrate robust performance on both scalp and intracranial EEG from the CHB-MIT and Freiburg datasets using a leave-one-seizure-out strategy for edge training. Across datasets, our approach achieved commendable and superior performance to state-of-the-art neural networks in metrics as AUROC, Sensitivity, and False Positive Rates (FPR), while drastically reducing memory footprint through quantization and lowering energy consumption by orders of magnitude. This work establishes a foundation for next-generation personalized Edge-AI systems capable of adapting and learning directly at the edge.