A linear regression model predicted carotid-femoral pulse wave velocity directly from raw PPG signals with a mean absolute error of 1.52 m/s and an R2 of 0.57.
Can carotid-femoral pulse wave velocity be accurately predicted from raw PPG signals using a machine learning-based linear regression model?
A simple linear regression model can predict carotid-femoral pulse wave velocity from raw PPG signals with moderate correlation, offering a potential non-invasive alternative to specialized instrumentation.
Effect estimate: MAE 1.52 m/s, RMSE 1.98 m/s, R2 0.57
Objective: Carotid-femoral pulse wave velocity (cfPWV) is a gold standard marker of arterial stiffness and cardiovascular risk. However, its traditional measurement requires specialized instrumentation. This study proposes a fully non-invasive, signal-based approach to predict cfPWV from photoplethysmography (PPG) signals by leveraging ensemble averaging and heart rate normalization. Design and method: A dataset of 283 raw PPG recordings sampled at 1000 Hz was used. The signals were segmented into cardiac cycles. Each cycle was resampled to 60 points and aligned by peak location. An ensemble average waveform was computed per recording, and heart rate (HR) was estimated from RR intervals, normalized between 0 and 1. The final feature vector per signal consisted of 61 values (60 waveform points + 1 HR value). A linear regression model was trained using 227 PPG ensemble average signals and evaluated on 56 independent recordings. Model training and prediction were implemented in MATLAB. Results: The trained model showed promising performance on the test set, with a mean absolute error (MAE) of 1.52 m/s and a root mean squared error (RMSE) of 1.98 m/s. The coefficient of determination (R2) was 0.57, indicating moderate but meaningful correlation between predicted and measured cfPWV. Conclusions: This study demonstrates the feasibility of predicting cfPWV directly from raw PPG signals using a simple linear model enriched by heart rate normalization. Future work will explore non-linear models and validation on larger, more diverse populations.
Obeid et al. (2026) studied Arterial stiffness (n=283). Linear regression model using PPG signals vs. Traditional measurement was evaluated on Prediction of carotid-femoral pulse wave velocity (cfPWV) (MAE 1.52 m/s, RMSE 1.98 m/s, R2 0.57). A linear regression model predicted carotid-femoral pulse wave velocity directly from raw PPG signals with a mean absolute error of 1.52 m/s and an R2 of 0.57.