A High-Accuracy and Ultra-Energy-Efficient Cardiac Arrhythmia Classification Processor for Wearable Intelligent ECG Monitoring

Wearable intelligent electrocardiography (ECG) sensors with integrated cardiac arrhythmia classification processors have been used to detect and classify arrhythmia, alerting users to potential cardiac diseases. While state-of-the-art arrhythmia classification processors employ neural networks (NNs), the high computational complexity of NNs results in significant energy consumption, limiting the model size and classification performance of NNs. Additionally, inter-patient variation in ECG can lead to accuracy degradation when applying a trained NN to patients whose ECG features differ from those in the training dataset. In this work, we propose an ultra-energy-efficient cardiac arrhythmia classification processor incorporating three key technologies: 1) heartbeat difference-based classification to improve accuracy under inter-patient variation and reduce energy consumption; 2) event-driven NN computation with shared feature extraction to reduce energy consumption; and 3) an adaptive NN wake-up technique to reduce energy consumption while maintaining accuracy. The design was fabricated using 55-nm CMOS process technology and evaluated using the MIT-BIH arrhythmia dataset. For arrhythmia classification, it demonstrates an energy consumption of 0.09 mu J per classification with 98.7%/96.6% accuracy for intra-patient and inter-patient testing, respectively.