Causal Inference in Cuffless Blood Pressure Estimation: A Pilot Study

Although photoplethysmogram (PPG) and electrocardiogram (ECG) signals have been used to estimate cuffless and continuous blood pressure (BP) for decades, most of the current popular methods are based on the correlated relationship between extracted features and BP. Current methods ignore causality in the system and lead to the unsatisfactory performance for BP estimation. This paper aims to infer the key features that cause BP changes and explore the feasibility of combining causal association with BP estimation problem. In the process, a total of 222 features extracted from PPG and ECG waveforms are used to infer causality with systolic BP (SBP) and diastolic BP (DBP) through fast causal inference (FCI) algorithm. The obtained causal graph suggests that the feature AMPPG(PPGvalley-sdPPGd) is the effect of SBP and AMPPG(PPGvalley-sdPPGb) is the effect of DBP, where AMPPG refers to the amplitude difference of PPG signal between two fiducial points and sdPPG is the second derivative of PPG signal. Moreover, the result provides new insights on features of amplitude class, in addition to the commonly studied pulse transit time (PTT). Inspired by Granger causality, time-lagged causal links are used to bridge the gap between causal graph and BP estimation and a causality-based multiple linear regression model for cuffless BP estimation is built. Compared with the corresponding correlation-based model, causality-based regression model achieves better performance for BP estimation, with mean error (ME) being 1.58±12.02, -4.67±9.03 mmHg and mean absolute difference (MAD) being 9.51, 7.54 mmHg for SBP and DBP, respectively.