Deep generative model with domain adversarial training for predicting arterial blood pressure waveform from photoplethysmogram signal

Keke Qin, Wu Huang, Tao Zhang

Abstract:

Background and Motivations: Continuous blood pressure (BP) monitoring is of critical importance to health state tracking and disease prevention. However, current mainstream BP measurement approaches are cuff-based, which is inconvenient and limit its usage scenarios. Predicting arterial blood pressure (ABP) could provide richer information than isolated BP values. The individual differences among data may hinder training.

Methods: A novel continuous, non-invasive and cuff-less approach is presented for generating ABP waveform using only raw photoplethysmogram (PPG) signal, from a signal conversion perspective, where a convolution-based deep autoencoder (DAE) model is developed. To overcome individual differences, Multi-domain adversarial training is merged with DAE (abbr. RDAE) to learn cross-domain features, and partial data is further used to calibrate (optional) the general model.

Results: The mean absolute error (MAE) of uncalibrated RDAE reached 7.945, 4.114 and 3.834 mmHg in systolic BP (SBP), diastolic BP (DBP) and mean BP (MBP) prediction. After using 80 s data for calibration, the MAE of RDAE reduced to 5.424, 3.144 and 2.885 mmHg accordingly.

Conclusion: Owning to the high-quality converted ABP segments, the resulting estimated BP is accurate. According to the BHS standard, RDAE achieved Grade C, Grade A and Grade A for SBP, DBP and MBP prediction, and the calibrated RDAE achieved Grade B, Grade A, Grade A accordingly.

Significance: Both domain adversarial training and calibration improve the performance in varying degrees. RDAE is competitive to other mainstream regression-based deep learning methods, while with fewer model parameters, and to other representative machine learning methods, while no need of complicated feature engineering.