A Reconfigurable 0.69-1.02nJ/Classification Biomedical AI Processor for Intelligent Health Monitoring Devices

Ultra-low-power biomedical AI processors are crucial for effective intelligent health monitoring systems (HMSs) in wearable devices. Traditional HMSs rely on feature-based methods [1], [2] that necessitate human intervention for pre-extract features before classification. Unfortunately, this approach often leads to poor accuracy <90%) and can be time-consuming. Recently, end-to-end neural networks (NNs) [3]–[8] have demonstrated promising results by directly processing original ExG (EEG, ECG, EMG, etc.) signals for classification. This method circumvents the limitations imposed by human expertise and achieves higher accuracy. Yet, the implementation of more complex NNs brings forth new challenges, as illustrated in Fig. 1: 1) prior reconfigurable HMSs underutilize memory resources (~8.9% utilization in the EEG mode of [7]) due to the need for multiple NNs of varying sizes for different detection modes; 2) frequent memory access significantly dominates the power consumption of AI systems; 3) the presence of redundant bio-information in ExG signals results in unnecessary computational energy dissipation.