Contactless Blood Pressure Measurement via Remote Photoplethysmography with Synthetic Data Generation Using Generative Adversarial Network

Deriving blood pressure in a non-invasive way via photoplethysmography (PPG) signals has become a familiar topic. With the knowledge of the relation between PPG and blood pressure, we expect to further make the measurement contactless for convenience reasons. An alternative signal source is remote photoplethysmography (rPPG) signals. There are mainly two kinds of approaches for exploiting blood pressure through PPG signals, one is by calculating the pulse transit time of the arterial pulse wave at two consecutive sites and the other is based on waveform feature analysis from a single signal. The calibration procedure is necessary for the former way, which leads to some limitations in general use. On the other hand, the properties of the rPPG waveform are far from PPG signals. Hence, the known waveform features in PPG signals are hard to be leveraged in the case of rPPG signals. Recently, convolutional neural networks are also applied for solving this problem. However, the lack of data is an obstacle to the training procedure and evaluation. In this study, a multichannel rPPG-based blood pressure estimator is proposed. To ease the data scarcity issue, the generative adversarial network is adopted to augment synthetic waveform data. Besides, as we know that some physiological states like age and BMI are dominant factors in blood pressure. InfoGAN is chosen in this work to generate the synthetic data with the blood pressure value fluctuating correspondingly to the controlled age and BMI combination. The proposed model outperforms the state-of-the-art methods on MIMIC III and Cuffless datasets. With the synthetic data generation, the mean absolute error (MAE) is reduced to 6.72 and 5.95 mmHg in MAP and DBP respectively. The standard deviations of the MAEs are also reduced. In the rPPG case, the MAE of SBP is 9.13 and 8.76 mmHg for DBP.