Table 5.

Models used to assess indicators of vascular age from PPG-derived parameters

Biophysical models

• Biophysical models: mostly use the Moens–Korteweg or Hughes equations to relate the vessel wall elastic modulus to PWV and distending pressure, respectively (208), to estimate systolic (SBP), diastolic (DBP) and pulse pressure (PP) from PPG-derived parameters. Regression models are used to estimate BPs from PAT or PTT (163) (e.g., linear, logarithmic, inverse square, or inverse).  • Models can account for additional factors: such as the pulsatile change in blood vessel diameter (estimated as PPG intensity ratio) (146), and the viscous effects of blood flow (55).  • Windkessel modeling: use of a Windkessel model to assess arterial compliance (103).

Statistical models

• Auto-regressive models: (with exogenous input—ARX, and moving-average—ARMA) have been used to estimate the central BP waveform (141) and arterial SBP and DBP (174) from a PPG signal.  • Estimating measures of vascular age: various regression models (e.g., linear, inverse, quadratic, exponential, partial least-squares) are used to estimate BPs from single PPG features (44, 59, 67, 79).

Machine learning (ML) models: estimation

• Estimating measures of vascular age from a single PPG pulse wave: the pulse wave, its first and second derivatives are used as inputs to a ML algorithm [e.g., nonlinear regression (39), deep neural network (41), support vector machine (SVM) (46)] to estimate numerical values (e.g., BPs).  • Estimating measures of vascular age from pulse wave features: features are used as inputs to a ML algorithm [e.g., AdaBoost (66), random forest (35), artificial neural network (ANN) (150), regression tree (175)] to estimate numerical values (e.g., BPs).  • Estimating measures of vascular age from multiple PPGs: features derived from multisite PPGs are used as inputs to ML algorithms (e.g., SVM) to estimate numerical values [e.g., SBP, DBP (61), and ABI (76)].  • Estimating measures of vascular age from PPG and other signals: PAT and other time and complexity features from the electrocardiogram and PPG, and PPG-derived features, are used as inputs to a ML algorithm (e.g., regularized linear regression, multiadaptive regression, back-propagation error neural network, convolutional neural network (CNN), SVM) to estimate numeric values, e.g., arterial DBP and SBP (31, 34, 45, 54, 67, 72). Associations with chronological age have also been assessed (47).  • Estimating measures of vascular age from PPG and demographics: use of time-, frequency-domain and statistical features of PPGs along with demographic data as an input to a ML algorithm (e.g., ensemble trees, Gaussian process regression, multiple linear regression) to estimate numeric values, e.g., SBP and DBP (18, 69).

Machine learning (ML) models: classification

• Classifying pulse waves: use of a ML algorithm [e.g., K-nearest neighbor (KNN), CNN] to classify a pulse wave or a PPG signal transformation into a diagnostic category, e.g., normo-, prehyper- and hyper-tension (28, 74).  • Classifying sets of pulse wave features: use of a ML algorithm [e.g., SVM, ANN, decision trees or KNN] to classify a set of pulse wave features into a diagnostic category, e.g., low or high PWV (105), normal or abnormal BP (56), normo-, prehyper- and hyper-tension (21, 75).  • Classifying and then estimating measures of vascular age based on category: use of two-step ML algorithms to classify PPG features into BP categories (e.g., using KNN) and then estimate numeric values (e.g., SBP and DBP estimated using regression trees optimized for each BP category) (40).

Machine learning (ML) models: miscellaneous

• Extracting features and estimating measures of vascular age from single PPG: use of ML algorithm (e.g., CNN) to extract morphological features from a PPG segment (64) or its spectrogram (70) to estimate numerical values (e.g., BPs).  • Improving the assessment of vascular aging: use of long short-term memory networks to capture temporal dependencies between PPG features (e.g., extracted by CNN) to better track changes in measures of vascular aging (e.g., BP) (17, 64, 77).  • Reducing the feature vector: use of a ML algorithm (e.g., ANN) to nonlinearly map PPG features to reduce feature vector before estimating measures of vascular age (e.g., BPs) (50).  • Reconstructing other signals: use of wavelet neural network (149) or auto-regressive model (141) to estimate BP waveform from PPG waveform.

PAT, pulse arrival time; PPG, photoplethysmogram; PTT, pulse transit time.