Scholarlink

<p>Cardiovascular screening on mobile and edge devices requires low-latency and low-power heart sound classification, yet many existing solutions are designed for general-purpose processors and graphics processors. This paper presents an end-to-end heterogeneous design that integrates mel-frequency spectral coefficient feature extraction and a lightweight convolutional neural network classifier on a Xilinx Zynq-7020 system-on-chip platform. The processing system coordinates data movement and control, while the programmable logic implements a streaming pipeline generated and optimized using high-level synthesis. The proposed implementation is evaluated on the PhysioNet and Computing in Cardiology Challenge 2016 dataset and achieves an accuracy of 90.1 percent, a precision of 88.3 percent, a recall of 93.9 percent, an F1 score of 0.911 and an area under the receiver operating characteristic curve of 96.07 percent. On the PYNQ-Z2 development board, the on-board system delivers an end-to-end latency of 4.6&#xa0;ms per segment, corresponding to approximately 217 segments per second, which provides a 6.4 times speedup compared with a central processing unit workflow of 29.4&#xa0;ms per segment. Post-synthesis results report resource usage of 46,881 look-up tables (88 percent), 28,137 flip-flops (26 percent), 132 block random access memory blocks (47 percent) and 188 digital signal processing slices (85 percent), with an estimated total platform power of approximately 2.28 watts. These results demonstrate real-time, energy-efficient heart sound classification on a resource-constrained heterogeneous field-programmable gate array platform.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An accelerated heart sound classification design based on a heterogeneous platform

  • 期刊论文

  • 发表日期：

    2026年3月3日

查看全文
  • Rongguo Yan,
  • Xiyun Zeng,
  • Yunhao Hu,
  • Qi Wang

摘要

Cardiovascular screening on mobile and edge devices requires low-latency and low-power heart sound classification, yet many existing solutions are designed for general-purpose processors and graphics processors. This paper presents an end-to-end heterogeneous design that integrates mel-frequency spectral coefficient feature extraction and a lightweight convolutional neural network classifier on a Xilinx Zynq-7020 system-on-chip platform. The processing system coordinates data movement and control, while the programmable logic implements a streaming pipeline generated and optimized using high-level synthesis. The proposed implementation is evaluated on the PhysioNet and Computing in Cardiology Challenge 2016 dataset and achieves an accuracy of 90.1 percent, a precision of 88.3 percent, a recall of 93.9 percent, an F1 score of 0.911 and an area under the receiver operating characteristic curve of 96.07 percent. On the PYNQ-Z2 development board, the on-board system delivers an end-to-end latency of 4.6 ms per segment, corresponding to approximately 217 segments per second, which provides a 6.4 times speedup compared with a central processing unit workflow of 29.4 ms per segment. Post-synthesis results report resource usage of 46,881 look-up tables (88 percent), 28,137 flip-flops (26 percent), 132 block random access memory blocks (47 percent) and 188 digital signal processing slices (85 percent), with an estimated total platform power of approximately 2.28 watts. These results demonstrate real-time, energy-efficient heart sound classification on a resource-constrained heterogeneous field-programmable gate array platform.