This is an openaccess article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Medical Informatics, is properly cited. The complete bibliographic information, a link to the original publication on https://medinform.jmir.org/, as well as this copyright and license information must be included.
For the noninvasive assessment of arterial stiffness, a wellknown indicator of arterial aging, various features based on the photoplethysmogram and regression methods have been proposed. However, whether because of the existing characteristics not accurately reflecting the characteristics of the incident and reflected waveforms of the photoplethysmogram or because of the lack of expressive power of the regression model, a reliable arterial stiffness assessment technique based on a single photoplethysmogram has not yet been proposed.
The purpose of this study is to discover highly correlated features from the incident and reflected waves decomposed from a photoplethysmogram waveform and to develop an artificial neural networkbased regression model for the assessment of vascular aging using newly derived features.
We obtained photoplethysmograms from 757 participants. All recorded photoplethysmograms were segmented for each beat, and each waveform was decomposed into incident and reflected waves by the Gaussian mixture model. The 26 basic features and 52 combined features were defined from the morphological characteristics of the incident and reflected waves. The regression model of the artificial neural network was developed using the defined features.
In correlation analysis, the features from the amplitude of the reflected wave and the skewness of the photoplethysmogram showed a relatively strong correlation with the participant’s real age. In the estimation of real age, the artificial neural network model showed 10.0 years of root mean square error. Its estimated age and real age had a strong correlation of 0.63 (
This study proved that the features defined from the reflected wave and skewness of the photoplethysmogram are useful to assess vascular aging. Moreover, the regression model of artificial neural network using these features shows the feasibility for the estimation of vascular aging.
Arterial stiffness is one of the major factors to clinically assess the risk of cardiovascular disease [
Photoplethysmogram (PPG), which is a noninvasive optical measuring technique of blood volume changes in microvessels, was also used to assess arterial stiffness. In the PPG waveform, the systolic phase and the diastolic phase repeatedly appear, corresponding to the cardiac systole and the cardiac diastole. The systolic phase indicates an increase in vascular blood volume, and the diastolic phase indicates a decrease in vascular blood volume [
In recent studies, machine learning techniques have been introduced to evaluate arterial stiffness. Dall’Olio et al [
The purpose of this study is to develop a new vascular aging assessment model using the PPG, which could be noninvasively and easily measured in daily life. In particular, unlike the existing PPGbased vascular aging estimation studies, we decompose the incident and reflected waves of the PPG waveform. New highly correlated features are then explored for vascular aging assessment from the decomposed PPG waves. Lastly, an ANNbased regression model with excellent nonlinear estimation performance is applied to estimate vascular aging.
Data were obtained from a total of 1000 patients who were scheduled for elective surgery (thyroid, breast, or abdominal) from July to September 2015 at Asan Medical Center. Through crosschecking of two researchers, 17 participants with loss of signal and 226 participants with indistinguishable PPG waveforms were excluded from the analysis. As a result, data from a total of 757 participants were used.
Characteristics of patients included in the analysis (N=757).
Category  Values  




Male  348 (46.0) 

Female  409 (54.0) 




PS 1  450 (59.4) 

PS 2  277 (36.6) 

PS 3  30 (4.0) 
Weight (kg), median (range)  61.8 (54.169.4)  
Height (cm), median (range)  161.6 (155.7168.0)  
BMI (kg/m^{2}), median (range)  23.5 (21.325.9)  




029  10 (1.3) 

3039  61 (8.1) 

4049  168 (22.2) 

5059  215 (28.4) 

6069  177 (23.4) 

7079  108 (14.3) 

8089  18 (2.4) 




Smoking  111 (14.7) 

Alcohol  240 (31.7) 




Hypertension  213 (28.1) 

Diabetes mellitus  90 (11.9) 

Pulmonary disease^{c}  15 (2.0) 

Renal disease^{d}  5 (0.7) 

Hepatic disease^{e}  23 (3.0) 

Neurologic disease^{f}  8 (1.1) 

Others^{g}  16 (2.1) 
^{a}ASA PS: American Society of Anesthesiologists Physical Status((1) a normal healthy patient, (2) a patient with mild systemic disease, and (3) a patient with severe systemic disease).
^{b}The median age is 56 years, with a range of 4665 years.
^{c}Pulmonary disease: asthma (7), emphysema (1), bronchiectasis (1), chronic obstructive pulmonary disease (5), and old tuberculosis (1).
^{d}Renal disease: chronic kidney disease (2) and end stage renal disease (3).
^{e}Hepatic disease: hepatitis B virus (11), hepatitis C virus (2), and liver cirrhosis (10).
^{f}Neurologic disease: stroke (1) and cardiovascular accident (7).
^{g}Others: angina (12), carotid artery stenosis (1), iron deficiency anemia (1), hyponatremia (1), and intracranial hemorrhage (1).
The measured signal was filtered using a finite impulse response bandpass filter having a 0.510 Hz passband, and then the pulse onset (ie, the start point of the waveform for each pulse) was detected (
Characteristics of the original PPG, incident and reflected waves, and reconstructed PPG for deriving candidate features. DIA: diastolic; INC: incident wave; INF: inflection point; OPPG: original photoplethysmogram; PPG: photoplethysmogram; REF: reflected wave; RPPG: reconstructed photoplethysmogram; SYS: systolic.
The features for vascular aging assessment consist of a basic feature defined from the specific points of the waveform before and after the decomposition of the incident and reflected waves of the PPG and a combined feature generated by combining the basic feature. Gaussian mixture model [
From the waveforms before and after the decomposition of the incident and reflected waves of the PPG, 26 basic features were generated for the development of the vascular aging estimation model.
Basic features defined from incident and reflected waves, first inflection point, reconstructed PPG^{a}, and original PPG.
Pulse type and feature  Definition  




Amplitude of incident wave’s peak 


Area of incident wave 


Time of incident wave’s peak 


Time period of incident wave 


Skewness of incident wave 


Kurtosis of incident wave 




Amplitude of reflected wave’s peak 


Area of reflected wave 


Time of reflected wave’s peak 


Time period of reflected wave 


Skewness of reflected wave 


Kurtosis of reflected wave 




Amplitude of first inflection point 


Time of first inflection point 


Area of first inflection 




Area of reconstructed PPG 


Skewness of reconstructed PPG 


Kurtosis of reconstructed PPG 




Amplitude of systolic peak 


Time of systolic peak 


Amplitude of diastolic peak 


Time of diastolic peak 


Area of original PPG 


Time period of original PPG 


Skewness of original PPG 


Kurtosis of original PPG 
^{a}PPG: photoplethysmogram.
^{b}INC: incident wave.
^{c}REF: reflected wave.
^{d}INF: inflection point.
^{e}RPPG: reconstructed photoplethysmogram.
^{f}SYS: systolic.
^{g}DIA: diastolic.
^{h}OPPG: original photoplethysmogram.
Domain and feature
In this study, since the actual age of participants is estimated based on various features extracted from their PPG, we used the ANN model, which is frequently used for nonlinear regression with independent features.
A leaveoneout crossvalidation (LOOCV) was used for the development and testing of the ANNbased regression model. In LOOCV, the entire data was divided into one test set and the rest assigned to the model development set. The model development set was divided into a training set and a validation set at a ratio of 8:2 with the same age distribution of participants. After training the model with the development set, the model was evaluated with the test set, and this process was repeated as many times as the number of data, so that all data were used for the model evaluation. The final performance of the model was obtained by averaging each evaluation result. The regression performance of the developed model was represented as RMSE. The ANNbased regression model proposed in this study was developed using 2.90 GHz Intel Core i710700 processor, 64 GB 1,333 MHz DDR4 RAM, NVIDIA Geforce RTX 2070 Super, Python 3.6.7: Anaconda, and Tensorflow 2.3.0.
Different values of hyperparameters for ANN^{a}based regression model for the estimation of vascular aging. Bold type indicates the hyperparameters for the optimal model.
Parameter  Value 
Input Layer Nodes 

Output Layer Nodes 

Hidden Layers Number  
Hidden Layer Nodes  64 
Activation Function 

Dropout Probability  0 0.1 0.3 
Kernel Initializer 

Loss Function 

Learning Rate  0.01 0.005 
Optimizer  SGD^{d} 
Early Stopping Patience  30 
Input Data Scaler  Standard 
^{a}ANN: artificial neural network.
^{b}ReLU: rectified linear unit.
^{c}MAE: mean absolute error.
^{d}SGD: stochastic gradient descent.
Architecture of the optimal version of the ANNbased regression model developed in this study. ANN: artificial neural network; INC: incident wave; OPPG: original photoplethysmogram; RPPG: reconstructed photoplethysmogram.
The Pearson correlation coefficient was calculated to investigate the relationship between the participants’ actual age and each feature, which was defined from the waveforms before and after the decomposition of PPG into the incident and reflected wave. The RMSE and coefficient of determination of the age estimated by the ANNbased vascular aging estimation model, which was developed with all the PPG features defined in this study, were calculated. In addition, using the estimated age and the actual age, a scatter plot and a BlandAltman plot were made and used to analyze the model's estimation performance.
The results of the correlation analysis between the actual age and the PPG features are as follows. The correlation coefficient between the actual age and the basic features, which is defined from the original PPG, the incident and reflected waves decomposed from PPG, and the reconstructed PPG, is shown in
Correlation coefficient and
Pulse type and feature  R^{b}  




0.06  


0.15  


0.23  


0.18  


–0.16  


–0.18  




–0.42  


–0.45  


0.10  


0.02  


0.19  


0.18  




–0.08  


0.18  


–0.04  




–0.39  


0.40  


0.04  




0.02  


0.27  


–0.39  


0.24  


0.06  


0.08  


0.41  


–0.07 
^{a}PPG: photoplethysmogram.
^{b}R: Pearson correlation coefficient.
^{c}INC: incident wave.
^{d}REF: reflected wave.
^{e}INF: inflection point.
^{f}RPPG: reconstructed photoplethysmogram.
^{g}SYS: systolic.
^{h}DIA: diastolic.
^{i}OPPG: original photoplethysmogram.
Correlation coefficient and
Domain and feature  R^{a}  



–0.18  

0.32  

0.32  

0.28  

0.38  

0.37  

0.34  

0.19  

–0.34  

–0.34  

0.31  

0.34  

–0.06  

0.42  



0.20  

0.15  

0.23  

0.22  

0.12  

0.24  

0.19  

0.16  

0.12  

–0.19  

0.03  

0.03  

0.03  



–0.28  

–0.22  

–0.11  

–0.02  

–0.09  

–0.15  

–0.11  

0.05  

0.02  

–0.02  

–0.36  

–0.32  

–0.22  

–0.17  

–0.22  

–0.40  

–0.36  

–0.27  

–0.25  

–0.28  

–0.33  

–0.28  

–0.17  

–0.10  

–0.16 
^{a}R: Pearson’s correlation coefficient.
^{b}INC: incident wave.
^{c}REF: reflected wave.
^{d}RPPG: reconstructed photoplethysmogram.
^{e}SYS: systolic.
^{f}OPPG: original photoplethysmogram.
The RMSE for the age estimation of the ANNbased regression model developed in this study was 10.0 years.
Scatter plot and coefficient of determination for the ANNbased regression model developed for the estimation of vascular aging in this study. ANN: artificial neural network.
BlandAltman plot for the ANNbased regression model developed for the estimation of vascular aging in this study. ANN: artificial neural network.
In this study, a highly correlated feature for assessing vascular aging was explored using features before and after decomposition of the incident and reflected waves of the PPG, and an ANNbased vascular aging estimation model was developed with the features derived. The ANNbased regression model showed the RMSE of 10.0 years in the age estimation. In comparing the correlation analysis before and after decomposition of the PPG incident and reflected waves, the feature defined after decomposition rather than before decomposition of the incident and reflected waves is useful for assessing vascular aging. In addition, in the comparison of all individual features, the feature defined from the reflected wave was confirmed as the best feature for assessing vascular aging. This reconfirms that changes in arterial stiffness due to vascular aging are reflected very well in the reflected wave characteristics of PPG, as Dawber et al [
For the model development, hyperparameter optimization, such as number of hidden layers, number of nodes, dropout rate, learning rate, optimizer, early stopping patience, and input data scaler of the ANNbased regression model, was performed. In determining the hidden layer, as the number of hidden layers and the number of nodes in the hidden layer decreased, the age estimation error of the proposed model tended to decrease. In addition, as the dropout ratio of the hidden layer increased, the estimation error decreased. This means that the proposed model has sufficient expressive power to overfit the training data and that performance can be improved by suppressing overfitting [
Comparison of the proposed model to the models of previous studies in root mean squared error, correlation coefficient, and
Reference and type of regression model  Input  RMSE^{a} (years)  R  
Proposed, ANN^{b}  Features from raw PPG^{c} and incident and reflected wave separated from raw PPG  10  0.63  
Millasseau et al [ 
Feature from raw PPG  N/A^{d}  –0.29  
Yousef et al [ 
Feature from raw PPG  N/A  –0.33  
Dall’Olio et al [ 
Raw PPG  12  N/A  N/A  
Chiarelli et al [ 






Linear  Feature from raw PPG and ECG^{f}  12  0.64  

ANN  Feature from raw PPG and ECG  11  0.74  

DCNN^{g}  Raw PPG and ECG  7  0.92 
^{a}RMSE: root mean squared error.
^{b}ANN: artificial neural network.
^{c}PPG: photoplethysmogram.
^{d}N/A: not applicable.
^{e}CNN: convolutional neural network.
^{f}ECG: electrocardiogram.
^{g}DCNN: deep convolutional neural network.
This study has some limitations. Most of the previous studies that performed vascular aging evaluation used finger PPG. However, in this study, vascular aging was evaluated based on nasal PPG. Therefore, it is difficult to generalize the results of this study to a vascular aging evaluation technique using PPG regardless of the measurement location. Therefore, it is necessary to analyze the agingrelated waveform change characteristics of PPG obtained from various measuring sites through additional studies. In addition, the ANNbased regression model developed in this study for estimating vascular aging is a relatively simple machine learning model with one hidden layer. Therefore, in future studies, it is necessary to improve the vascular aging estimation performance by applying a more sophisticated machine learning technique with increased model complexity. Moreover, this study did not investigate various risk factors that can accelerate vascular disease, such as atherosclerosis; therefore, it is necessary to evaluate the model performance and examine the possibility of application according to various subject characteristics.
In this study, we derived various features from the decomposed PPG waveforms before and after decomposition of the waveform into incident and reflected waves to explore features highly correlated with vascular aging, and it was confirmed that the reflected waverelated features had a strong correlation with participant’s age. In addition, the ANNbased regression model developed using the derived feature had 10 years of RMSE in estimating the participants’ actual age and showed the improved vascular aging estimation performance in comparison with the models introduced in previous studies. These results suggest that the developed technology can be applied to a wearable device and used to assess vascular health in reallife situations. However, this study was performed based on nasal PPG, not finger PPG, which is not frequently used in vascular aging evaluation studies. Since it is not clear whether the change in the PPG waveform due to vascular aging has a specific pattern for each measurement location, additional research needs to be performed for clarification.
augmentation index
artificial neural network
convolutional neural network
deep convolutional neural network
diastolic
electrocardiogram
incident wave
inflection point
leaveoneout crossvalidation
mean absolute error
original photoplethysmogram
photoplethysmogram
reflected wave
rectified linear unit
reflection index
root mean squared error
reconstructed photoplethysmogram
stiffness index
systolic
This research was supported by the Basic Science Research Program through the National Research Foundation of Korea (NRF), funded by the Ministry of Education (NRF 2018R1D1A3B07046442), Republic of Korea, and supported by the Korea Health Technology R&D Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health and Welfare, South Korea, (HI21C0011).
JP contributed to data analysis, drafting, writing, and figure design and drawing. HS contributed to the conception and design of the work, supervision, writing, and indepth review. All authors contributed to the critical review of the final document.
None declared.