BioWAP: A Reconfigurable Biomedical AI Processor with Adaptive Processing for Co-Optimized Accuracy and Energy Efficiency

Intelligent wearable/implantable health monitoring devices integrating biomedical AI processors have been developed for automatically identifying abnormality in users' biomedical signals. Three features are required for the biomedical AI processors, including high accuracy, low energy consumption and reconfigurability. However, the existing designs focus on achieving high energy efficiency which sacrifices accuracy and reconfigurability. To address these issues, in this work, a reconfigurable biomedical AI processor with diverse adaptive processing techniques has been proposed for co-optimized accuracy and energy-efficiency. The key features include 1) adaptive feature-fusion based classification architecture for improving the classification accuracy with low computation complexity. 2) adaptive-window based neural network processing architecture to improve both accuracy and energy efficiency. 3) K-Nearest-Neighbors (KNN) based adaptive weight precision selection technique to reduce the energy consumption while maintaining high accuracy. The proposed design is implemented and fabricated with 55nm CMOS process technology. Being highly reconfigurable, it achieves high accuracy (98.7%, 98.5% and 99.87%) and low energy (0.18 μJ, 2.3 μJ and 1.1 μJ) for three typical biomedical AI tasks (i.e. ECG arrhythmia classification, ECG atrial fibrillation detection and EEG seizure detection), outperforming the state-of-the-art designs.