Features from the photoplethysmogram and the electrocardiogram for estimating changes in blood pressure

Abstract There is increasing emphasis being placed on cuffless blood pressure (BP) monitoring using the electrocardiogram (ECG) and photoplethysmogram (PPG). These signals have previously been employed to compute estimates of the pulse arrival time (PAT, the time between characteristic fiducial points on the ECG and PPG) as a surrogate for BP. Current work in this field is focused on using either the ECG or the PPG alone as a single source BP estimator using characteristic morphological features and machine learning models. However, the appropriate features and models to use remain unclear. As a result, BP estimation using the ECG or the PPG signals alone has produced inconclusive results. In this work, we investigated the best features available from the PPG and the ECG for BP estimation using both linear and non-linear models. We conducted a clinical study involving 30 healthy volunteers (53.8% female, 28 (± 9) years old, with a body mass index of 22.5± (5.2 kg/m2). Each session lasted 28.0 (± 0.12) minutes and BP was varied by administering an infusion of phenylephrine, a medication that causes arterial and venous vasoconstriction. We extracted a large and diverse set of features from both the PPG and the ECG and assessed their individual importance for estimating changes in BP (∆BP) using a ranking coefficient. In addition to features commonly used in the literature, we propose new features extracted from both signals. We implemented linear (ordinary least squares, OLS) and non-linear (random forest, RF) machine learning models to estimate ∆BP. We adopted a hybrid calibration strategy by including patient demographics in the feature set. We trained, tuned, and evaluated these models in a nested leave-one-subject-out cross-validation framework and we reported the results as correlation coefficient (ρp), root mean squared error (RMSE), and mean absolute error (MAE). We compared our results to those of estimating ∆BP using PAT. The non-linear RF model significantly (p < 0.05) outperformed the linear OLS model using both the PPG and the ECG signals across all performance metrics. Estimating ∆SBP using the PPG alone (ρp = 0.86 (0.23), RMSE = 5.66 (4.76) mmHg, MAE =4.86 (4.29) mmHg) performed significantly better than using the ECG alone (ρp = 0.69 (0.45), RMSE = 6.79 (4.76) mmHg,MAE = 5.28 (4.57) mmHg), all p < 0.001. Estimating ∆BP using features from the PPG alone had a similar performance to that of using PAT (which requires a simultaneous ECG signal). Kurtosis of the PPG waveform showed consistently high feature ranking for both the OLS and RF models. Additionally, the highest ranking features from the PPG largely modelled increasing reflected wave interference driven by changes in arterial stiffness. This finding was supported by changes observed in the PPG waveform in response to the phenylephrine infusion. However, a large number of features were required for accurate BP estimation, highlighting the high complexity of the problem. We conclude that the PPG alone may be further explored as a potential single source, cuffless, blood pressure estimator. The use of the ECG alone is not justified. Non-linear models may perform better as they are able to incorporate interactions between feature values and demographics. However, demographics may not adequately account for the unique and individualised relationship between the extracted features and BP.