4.5 BioAIP: A Reconfigurable Biomedical AI Processor with Adaptive Learning for Versatile Intelligent Health Monitoring

Abstract

Intelligent health monitoring devices automatically detect abnormalities in users' biomedical signals (e.g. arrhythmia from an ECG signal or a seizure from an EEG signal) through signal classification. Compared to conventional machine learning methods, neural-network-based AI classification methods are promising in achieving higher classification accuracy, but with significantly increased computational complexity, posing challenges to real-time performance and low power consumption. AI processors have been designed to accelerate neural networks for general AI applications such as image and voice recognition. They are not suitable for biomedical AI processing, which requires a combination of biomedical and AI processing hardware. In addition, the design redundancy for general AI applications results in large power consumption making it unsuitable for ultra-low-power health monitoring devices. There are also some biomedical AI processors such as ECG/EEG/EMG AI processors. However, they are customized for specific algorithms and tasks, prohibiting algorithm upgrades, limiting their applicability. In addition, prior designs lack adaptive learning to address the patient-to-patient variation issue. In this work, the BioAlP is proposed - a reconfigurable biomedical Al processor with adaptive learning. It has the following key features: 1) A reconfigurable biomedical Al processing architecture with reconfigurable neural network and biomedical processing engines to support versatile biomedical Al processing. 2) An event-driven biomedical Al processing architecture and approximate data compression technique to reduce power consumption. 3) An Al-based adaptive-learning architecture to address patient-to-patient variation. 4) A reconfigurable FIR engine reusing the neural-network engine to reduce the hardware overhead.