E-NPU: A 34~126nJ/Class Event-Driven Adaptive Neural SoC with Signal-Dynamics-Aware Feature Clustering and Multi-Model In-Memory Inference/Training for Personalized Medical Wearables

Biomedical signal processing systems-on-chip (SoC) have shown significant promise in human-machine interfaces and closed-loop neuromodulation applications [1]–[5]. Future personalized medical wearables are anticipated to provide long-term monitoring and accurate interpretation while maintaining minimal energy consumption [6]–[8]. Lightweight neural networks, such as binary, spiking, and convolutional neural networks (BNN/SNN/CNN), are promising for neural signal analysis, with neural architecture search (NAS) techniques enabling the deployment of task-specific models optimized for accuracy and efficiency [9]–[12]. However, developing an accurate, ultra-energy-efficient medical SoC that seamlessly integrates signal feature extraction and processing with multi-model reconfigurability remains challenging. Furthermore, inter-patient variance adversely affects the accuracy of biosignal classifications in real-world scenarios, thereby limiting the reliability of these devices [13], [14].