Energy-Efficiency-Optimized Convolutional Spiking Neural Networks for Patient-Specific Seizure Detection

This thesis presents the design, development, and hardware implementation of neuromorphic spiking convolutional neural networks (SNNs) for patient-specific seizure detection. Leveraging the energy-efficient, event-driven nature of SNNs, the work achieves a low-power, low-latency, high-accuracy processing unit for multi-channel implantable brain-machine interfaces. Initial validation on publicly available datasets using spiking convolutional neural networks (SCNNs) demonstrates average sensitivities and specificities of 83.02% and 86.31%, respectively, with a false positive rate of 0.69/hour. The 1-bit Integer-Net SCNN achieves sensitivity comparable to 32-bit CNNs while improving memory efficiency by 27× and energy efficiency by 98.6%, consuming just 1.28 μJ/classification. Further optimization tailored to patient-specific seizure detection on STM32H747 microcontrollers and Intel Loihi simulators achieves sensitivities of 92.2% and specificities of 97.3%. By optimizing quantization resolution, encoding schemes, and implementing SCNNs in HDL, this thesis advances energy-efficient seizure detection, contributing to adaptive brain interfaces and the broader field of neuromorphic computing.