Bian Dayi, Mehta Pooja, Selvaraj Nandakumar
Annu Int Conf IEEE Eng Med Biol Soc. 2020 Jul;2020:5948-5952. doi: 10.1109/EMBC44109.2020.9176231.
Respiratory rate (RR) is an important vital sign marker of health, and it is often neglected due to a lack of unobtrusive sensors for objective and convenient measurement. The respiratory modulations present in simple photoplethysmogram (PPG) have been useful to derive RR using signal processing, waveform fiducial markers, and hand-crafted rules. An end- to-end deep learning approach based on residual network (ResNet) architecture is proposed to estimate RR using PPG. This approach takes time-series PPG data as input, learns the rules through the training process that involved an additional synthetic PPG dataset generated to overcome the insufficient data problem of deep learning, and provides RR estimation as outputs. The inclusion of a synthetic dataset for training improved the performance of the deep learning model by 34%. The final mean absolute error performance of the deep learning approach for RR estimation was 2.5±0.6 brpm using 5-fold cross-validation in two widely used public PPG datasets (n=95) with reliable RR references. The deep learning model achieved comparable performance to that of a classical method, which was also implemented for comparison. With large real-world data and reference ground truth, deep learning can be valuable for RR or other vital sign monitoring using PPG and other physiological signals.
呼吸频率（RR）是健康状况的一项重要生命体征指标，但由于缺乏用于客观便捷测量的非侵入式传感器，它常常被忽视。简单光电容积脉搏波描记图（PPG）中呈现的呼吸调制已被证明有助于通过信号处理、波形基准标记和手工规则来推导RR。本文提出了一种基于残差网络（ResNet）架构的端到端深度学习方法，用于利用PPG估计RR。该方法将时间序列PPG数据作为输入，通过训练过程学习规则，训练过程中涉及生成一个额外的合成PPG数据集以克服深度学习的数据不足问题，并将RR估计作为输出。包含合成数据集进行训练使深度学习模型的性能提高了34%。在两个广泛使用的具有可靠RR参考值的公共PPG数据集（n = 95）中，使用5折交叉验证，深度学习方法用于RR估计的最终平均绝对误差性能为2.5±0.6次/分钟。深度学习模型实现了与经典方法相当的性能，经典方法也被实现用于比较。对于使用PPG和其他生理信号进行RR或其他生命体征监测而言，深度学习在拥有大量真实世界数据和参考真值的情况下可能具有重要价值。
