April 13, 2021BiosensorsOpen Access

Using CNN and HHT to Predict Blood Pressure Level Based on Photoplethysmography and Its Derivatives

View Full Paper

Ask AI

Bookmark

Why the study?

Traditional cuff measurements do not adequately represent blood pressure over time, creating an urgent need for portable, continuous, easy-to-operate, and low-cost blood pressure measuring devices.

Can a CNN based on the HHT method using PPG and its derivatives accurately predict blood pressure risk levels?

Structured PICO

Can a CNN based on the HHT method using PPG and its derivatives accurately predict blood pressure risk levels?

Population

582 data records (10 s each) from the Medical Information Mart for Intensive Care (MIMIC) database

Intervention

Convolutional neural network (CNN) based on the Hilbert-Huang Transform (HHT) method using photoplethysmography (PPG) and its derivatives (PPG+ dataset)

Outcome

Prediction of blood pressure risk level (normotension vs hypertension, normotension vs prehypertension, and normotension+prehypertension vs hypertension)

A CNN model using Hilbert-Huang Transform of PPG signals and its derivatives can accurately predict blood pressure risk levels, offering a potential method for continuous, non-invasive BP monitoring.

Implication

May aid wearable BP monitoring; leaves open prospective validation before clinical use.

Abstract

According to the WTO, there were 1.13 billion hypertension patients worldwide in 2015. The WTO encouraged people to check the blood pressure regularly because a large amount of patients do not have any symptoms. However, traditional cuff measurement results are not enough to represent the patient's blood pressure status over a period of time. Therefore, there is an urgent need for portable, easy to operate, continuous measurement, and low-cost blood pressure measuring devices. In this paper, we adopted the convolutional neural network (CNN), based on the Hilbert-Huang Transform (HHT) method, to predict blood pressure (BP) risk level using photoplethysmography (PPG). Considering that the PPG's first and second derivative signals are related to atherosclerosis and vascular elasticity, we created a dataset called PPG+; the images of PPG+ carry information on PPG and its derivatives. We built three classification experiments by collecting 582 data records (the length of each record is 10 s) from the Medical Information Mart for Intensive Care (MIMIC) database: NT (normotension) vs. HT (hypertension), NT vs. PHT (prehypertension), and (NT + PHT) vs. HT; the F1 scores of the PPG + experiments using AlexNet were 98.90%, 85.80%, and 93.54%, respectively. We found that, first, the dataset established by the HHT method performed well in the BP grade prediction experiment. Second, because the Hilbert spectra of the PPG are simple and periodic, AlexNet, which has only 8 layers, got better results. More layers instead increased the cost and difficulty of training.