Blood Pressure Prediction Using Information Entropy of Electrocardiogram and Photoplethysmogram Signals

Hong Hao^1*, Kevin Sheng-Kai Ma², Tzay-Ming Hong¹

¹Department of Physics, National Tsing Hua University, Hsinchu 30043, Taiwan
²Center for Global Health, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA

* Presenter:Hong Hao, email:physicsred@gmail.com

It is well known that hypertension cannot be detected by symptom at the early stage. Furthermore, the traditional method to monitor blood pressure (BP) cannot be applied anytime and anywhere. Hence, it is necessary to develop wearable devices to supervise BP. Entropies can be used to quantify information, and some of them have been derived from biomedical signals. In this work, we emphasize the importance of Shannon entropy, sample entropy, and permutation entropy. Derived from electrocardiogram (ECG) /photoplethysmogram (PPG) data, these entropies are promising features to combine with machine learning (Linear regression (LR), Random forest (RF), support vector regression (SVR), deep neural network (DNN), XGBoost to enhance BP prediction performance (Root Mean Squared Error (RMSE), accuracy). We found that (1) prediction of systolic blood pressure (SBP) is best to incorporate permutation entropy, while sample entropy works best for diastolic blood pressure (DBP); (2) PPG (/ ECG+PPG) entropy promotes performance for SBP and DBP (SVR, LR, RF, DNN, and XGBoost); (3) the performance of SVR and DNN is better with entropic features; (4) PPG entropy is better than ECG entropy. In summary, this work indicates that entropies have the potential to be used as features for machine learning in wearable devices to monitor BP.

Keywords: Entropy, electrocardiogram, photoplethysmogram, blood pressure