Efficient EEG Seizure Detection Using INT8 Quantization, Channel Pruning, and Spiking Neural Networks

This paper evaluates three efficiency strategies for continuous EEG seizure detection on wearable or implantable devices, using a single 1D CNN on the CHB-MIT scalp EEG dataset as a common baseline. The authors convert the CNN to an SNN through parameter transfer, prune EEG channels alongside 2:4 structured weight sparsity, and apply INT8 quantization through FX- and ONNX-based workflows with quantization-aware training and operator fusion. Quantized variants reduce model storage from 1.63 MB to 0.44 MB, cut estimated energy per inference by up to 64%, and improve CPU latency by up to 2.8x while preserving or slightly improving AUC.