Entropy-Facilitated Machine Learning for Blood Pressure Estimation Using Electrocardiogram and Photoplethysmogram in a Wearable Device

Wearable devices that detect electrocardiogram (ECG) signals and photoplethysmogram (PPG) signals have been proposed to be effective at identifying early stage hypertension and estimating blood pressure. On the other hand, information entropy has been applied to determine whether certain biomedical signals represent pathological changes. In this study, Shannon entropy, sample entropy, and permutation entropy were derived from ECG and PPG signals collected through a wrist band wearable device, and were used to test whether the combination use of common features extracted from ECG and PPG, plus information entropies of ECG and PPG, may serve as effective features for blood pressure estimation when using machine learning-based linear regression (LR), random forest (RF), support vector regression (SVR), deep neural network (DNN), and XGBoost. Overall, the accuracy for blood pressure estimation was higher for diastolic blood pressure (DBP) than that for systolic blood pressure (SBP). The use of the entropies of ECG, PPG, or both, may increase the performance of BP estimation at an increase ranging from 3.3% to 10%. Accuracy of DBP estimation reached the highest when using entropies of ECG and PPG in either SVR or RF, with RF having a lower root-mean-square error (RMSE) compared with that for SVR. Likewise, SVR outperformed other models for the estimation of systolic blood pressure (SBP). The use of PPG entropy benefited the performance of LR, RF, and DNN in SBP estimation, which was better than when using entropies of ECG; for DNN, PPG entropy also brought about higher accuracy when it comes to the estimation of DBP. In conclusion, the use of entropies of ECG and PPG can improve the performance of blood pressure estimation, thus appears to be useful features in wearable devices that may facilitate blood pressure monitoring.