Customized Development and Hardware Optimization of a Fully-Spiking SNN for EEG-Based Seizure Detection

We present a customized approach in development of a fully-spiking neural networks (SNNs) tailored for energy-efficient data-driven signal processing in implantable brain neural interfaces. The importance of a customized design lies in its ability to optimize hardware and energy efficiency while maintaining high classification performance. Our approach allows for the customization of key parameters, including quantization resolution of weights and biases, encoding scheme, encoder placement, temporal resolution, neuron type, and internal parameters such as threshold value, resetting process, and refractory period. We demonstrated the efficacy of this customization in improving hardware and energy efficiency through model development, software-based training and testing, and subsequent synthesis using Verilog RTL on FPGAs and ASIC implementation. Performance evaluation using the CHB-MIT dataset showed an average sensitivity of $\mathbf{9 2. 2 \%}$ and specificity of $\mathbf{9 7. 3 \%}$ for seizure detection. The synthesis reports provided insights into the memory, computation, and energy requirements for hardware implementation, highlighting the efficiency and effectiveness of our approach. Our results show that SNN models leads to only 1% drop of sensitivity compared with a 32-bit real-value resolution SCNN model, while offering more than 4 times improvement in memory efficiency.