Wrist-worn photoplethysmography (PPG) sensors present a popular alternative to electrocardiogram recording for heart rate (HR) estimation. However, their accuracy is limited by motion artifacts inherent in ambulatory settings. In this paper, we propose a binarized neural network framework, b-CorNET, to efficiently estimate HR from single-channel wrist PPG signals during intense physical activity. The model comprises two binary convolution neural network layers followed by two binary long short-Term memory (b-LSTM) layers and a dense layer working on quantized PPG data. The proposed framework achieves an MAE of 3.75±3.05 bpm when evaluated on 12 IEEE SPC subjects. Furthermore, a novel, low-complexity architecture for the b-LSTM layers is proposed and efficiently mapped on a Xilinx Virtex5 FPGA, enabling HR computation.
|Original language||English (US)|
|Title of host publication||BioCAS 2019 - Biomedical Circuits and Systems Conference, Proceedings|
|Publisher||Institute of Electrical and Electronics Engineers Inc.|
|State||Published - Oct 2019|
|Event||2019 IEEE Biomedical Circuits and Systems Conference, BioCAS 2019 - Nara, Japan|
Duration: Oct 17 2019 → Oct 19 2019
|Name||BioCAS 2019 - Biomedical Circuits and Systems Conference, Proceedings|
|Conference||2019 IEEE Biomedical Circuits and Systems Conference, BioCAS 2019|
|Period||10/17/19 → 10/19/19|
Bibliographical noteFunding Information:
V. CONCLUSION In this paper a b-CorNET framework with a proof-of-concept b-LSTM implementation is presented to estimate HR from wrist PPG data. The results in comparison to nonbinary CorNET show a slight degradation (< 0.1bpm), due to binarization and quantization. Future work will further increase the robustness of the model and integrate the CNN layers along with LSTM for a complete PPG HR estimation system. ACKNOWLEDGMENT The first author would like to thank IFRS, CAPES and CNPq Brazilian agencies for supporting this research. REFERENCES
© 2019 IEEE.
Copyright 2020 Elsevier B.V., All rights reserved.
- Binary LSTM