Scholarlink

<p>Respiratory disorders, particularly in patients with conditions such as COVID-19, pose significant health risks. Respiration rate (RR) is a critical vital sign that provides valuable insights into a patient’s health status. Abnormalities in RR often indicate health deterioration, making continuous monitoring an essential early warning system. However, RR monitors are typically limited to patients in the intensive care unit (ICU). Current studies have explored the potential of photoplethysmogram (PPG) signals for noninvasive RR estimation. Although several deep learning (DL) solutions for RR prediction have been developed, challenges remain in achieving high accuracy and real-time deployment. This research proposes a novel lightweight hybrid DL model that integrates SqueezeNet and a transformer with multi-scale attention fusion to enhance feature extraction from PPG signals for effective RR estimation. The model was evaluated using two widely recognized public datasets, Beth Israel Deaconess Medical Center (BIDMC) and CapnoBase. Validation metrics included mean absolute error (MAE), percentage error (E), root mean squared error (RMSE), correlation coefficient (R), standard deviation (2SD), and Pearson correlation coefficient (PCC). Experimental results demonstrate that the proposed approach outperforms four well-established DL-based methods and the most recent models in the literature. For the BIDMC dataset, the proposed model achieved a maximum R and PCC of 0.947 and 0.72, respectively, along with minimal error metrics: MAE of 0.576 breaths per minute (bpm), RMSE of 1.057&#xa0;bpm, E of 0.04, and 2SD of 1.86&#xa0;bpm. On the CapnoBase dataset, the model achieved the highest PCC of 0.78 and R of 0.962, with minimal error scores: MAE of 0.465&#xa0;bpm, RMSE of 0.897&#xa0;bpm, E of 0.03, and 2SD of 1.73&#xa0;bpm. An ablation study further validated the significance of each module within the architecture. To enable real-time deployment, the model was implemented on an Intel Arria 10 GX1150 FPGA, optimizing computational efficiency, reducing power consumption, and improving inference speed. This advancement supports broader telemedicine and wearable healthcare applications, with potential for further enhancements through model scaling and multimodal signal integration.</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

FPGA-based SqueezeNet transformer with multi-scale attention fusion for real-time respiration rate estimation using PPG signals

  • 期刊论文

  • 发表日期：

    2025年6月18日

查看全文
  • Immaculate Joy Selvam,
  • Srinivasan Chandran,
  • Saravanakumar Chandrasekaran

摘要

Respiratory disorders, particularly in patients with conditions such as COVID-19, pose significant health risks. Respiration rate (RR) is a critical vital sign that provides valuable insights into a patient’s health status. Abnormalities in RR often indicate health deterioration, making continuous monitoring an essential early warning system. However, RR monitors are typically limited to patients in the intensive care unit (ICU). Current studies have explored the potential of photoplethysmogram (PPG) signals for noninvasive RR estimation. Although several deep learning (DL) solutions for RR prediction have been developed, challenges remain in achieving high accuracy and real-time deployment. This research proposes a novel lightweight hybrid DL model that integrates SqueezeNet and a transformer with multi-scale attention fusion to enhance feature extraction from PPG signals for effective RR estimation. The model was evaluated using two widely recognized public datasets, Beth Israel Deaconess Medical Center (BIDMC) and CapnoBase. Validation metrics included mean absolute error (MAE), percentage error (E), root mean squared error (RMSE), correlation coefficient (R), standard deviation (2SD), and Pearson correlation coefficient (PCC). Experimental results demonstrate that the proposed approach outperforms four well-established DL-based methods and the most recent models in the literature. For the BIDMC dataset, the proposed model achieved a maximum R and PCC of 0.947 and 0.72, respectively, along with minimal error metrics: MAE of 0.576 breaths per minute (bpm), RMSE of 1.057 bpm, E of 0.04, and 2SD of 1.86 bpm. On the CapnoBase dataset, the model achieved the highest PCC of 0.78 and R of 0.962, with minimal error scores: MAE of 0.465 bpm, RMSE of 0.897 bpm, E of 0.03, and 2SD of 1.73 bpm. An ablation study further validated the significance of each module within the architecture. To enable real-time deployment, the model was implemented on an Intel Arria 10 GX1150 FPGA, optimizing computational efficiency, reducing power consumption, and improving inference speed. This advancement supports broader telemedicine and wearable healthcare applications, with potential for further enhancements through model scaling and multimodal signal integration.