An Energy-Efficient ECG Processor With Ultra-Low-Parameter Multistage Neural Network and Optimized Power-of-Two...

Area of Science:

• Biomedical Engineering
• Computer Engineering
• Artificial Intelligence

Background:

• Cardiac arrhythmias pose a significant health risk, necessitating accurate and efficient detection methods.
• Existing electrocardiogram (ECG) processing solutions often face challenges with energy efficiency and computational complexity.
• Algorithm-hardware co-design offers a promising approach to optimize specialized hardware for complex signal processing tasks.

Purpose of the Study:

• To develop an energy-efficient ECG processor for real-time cardiac arrhythmia classification.
• To design a lightweight, two-stage neural network architecture optimized for hardware implementation.
• To improve low-bit quantization accuracy and reduce computational overhead in ECG processing.

Main Methods:

• Algorithm-hardware co-design integrating pre-processing and neural network acceleration.
• A novel two-stage neural network: Discrete Wavelet Transform (DWT) + ultra-low-parameter Multilayer Perceptron (MLP), followed by group convolution and channel shuffle.
• Optimized Power-of-Two (OPOT) quantization and a multiplier-less processing element for efficient low-bit operations.
• Implementation on a 65nm CMOS process with reconfigurable processing elements and adapted memory blocks.

Main Results:

• Achieved 98.59% accuracy for 5-class cardiac arrhythmia classification on the MIT-BIH dataset using 4-bit weight precision.
• Demonstrated an energy consumption of 0.15 uJ per inference at 1V.
• Realized a 64% energy saving compared to state-of-the-art approaches.
• Successfully implemented a lightweight neural network architecture with hardware resource optimization.

Conclusions:

• The proposed energy-efficient ECG processor effectively classifies cardiac arrhythmias with high accuracy and reduced power consumption.
• The algorithm-hardware co-design approach and novel neural network architecture are key to achieving significant energy savings.
• The OPOT quantization and multiplier-less design contribute to efficient low-bit precision processing for wearable or implantable cardiac monitoring devices.