Reconfigurable Low-Overhead SoC for Arrhythmia Detection Based on Cyclic Convolutional Neural Networks

Cardiovascular disease continue to pose a major threat to the lives and health of humans across the world. How to achieve portable and high-precision detection has become a research hotspot in preventing cardiovascular diseases. To address the issue of high overhead and difficulty in reconfigurable in high-precision artificial intelligence (AI) detection. This paper proposes a reconfigurable low-overhead arrhythmia detection system on chip (SoC) solution based on cyclic convolutional neural networks (CNN). The solution first involves researching the algorithms and structures of convolutional neural networks. This solution first studies the algorithms and structures of convolutional neural networks. Using cyclic CNN to classify arrhythmia, and reusing convolution and pooling modules to reduce hardware overhead. Then, by flexibly configuring the hyperparameter registers within the cyclic CNN-IP core, combined with the RISC-V instruction programming CPU method, achieved the reconfigurable CNN on SoC. Finally, in the digital backend clock gating and multi-threshold voltage techniques are adopted to reduce chip power consumption by 64.3%. In this paper, under the TSMC 65nm process the chip layout design is completed. With a supply voltage of 1V and an operating frequency of 100MHz, the classification accuracy for arrhythmias in the MIT-BIH dataset reaches 97.07%, with a chip area of 3.68 mm2 and a power consumption of 2.86 mW.