A fully-mapped and energy-efficient FPGA accelerator for dual-function AI-based analysis of ECG

Area of Science:

• Biomedical Engineering
• Computer Engineering
• Artificial Intelligence

Background:

• Artificial intelligence (AI) analysis of electrocardiogram (ECG) data is crucial for diagnosing cardiac conditions.
• Existing hardware accelerators often face limitations in performance and energy efficiency for real-time ECG analysis.
• Resource-constrained applications like wearable devices require optimized solutions for on-device ECG processing.

Purpose of the Study:

• To propose a fully-mapped Field Programmable Gate Array (FPGA) accelerator for AI-based ECG analysis.
• To develop a dual-function accelerator integrating a 1-D convolutional neural network (CNN) and a heart rate estimator.
• To optimize hardware implementation for high performance and energy efficiency in ECG monitoring.

Main Methods:

• A fully-mapped 1-D CNN is implemented on an Intel Cyclone V FPGA, projecting each layer to a hardware module.
• A virtual flatten layer is introduced to bridge feature extraction and fully-connected layers, maximizing parallelism.
• A dedicated, pipelined heart rate estimator with self-adaptive thresholding and optimized calculation is implemented on the FPGA.

Main Results:

• The FPGA accelerator achieved a 43.08x speedup for the 1-D CNN compared to an ARM processor and 8.38x vs. an Intel CPU.
• The heart rate estimator demonstrated 25.48x and 1.55x speedups over software implementations on the respective platforms.
• The accelerator achieved an energy efficiency of 63.48 GOPS/W with a power consumption of 67.74 mW.

Conclusions:

• The proposed fully-mapped FPGA accelerator offers significant performance improvements for AI-based ECG analysis.
• The dual-function design provides efficient and parallel processing for both CNN inference and heart rate estimation.
• The accelerator's high energy efficiency and low power consumption make it ideal for wearable and portable ECG monitoring devices.