Physiological Waveform Features Based Aging Clocks Associate with Cardiovascular Disease
- 1UCLA
Abstract
Aging affects the morphology of photoplethysmography (PPG) and arterial blood pressure (ABP) waveforms. Here, we demonstrate that machine learning (ML) models trained on features derived from PPG or ABP waveforms can predict an individual's age. We further investigated whether the discrepancy between predicted and chronological age (Age-Gap) could serve as a measure of cardiovascular health by testing its association with cardiovascular disease (CVD). From a large perioperative dataset of 17,327 surgical patients, high-fidelity PPG and ABP waveforms were analyzed, and over 840 time-, frequency-, and statistical-domain features per cardiac cycle were extracted using a signal-processing-based tool. These features were then used to develop ML-based aging clocks, PPGAge and ABPAge, trained on PPG and ABP features, respectively. For each aging clock, we evaluated five ML models: Light Gradient Boosting Machine (LightGBM), eXtreme Gradient Boosting (XGBoost), Random Forest, Lasso, and Elastic Net. Individuals were classified as having accelerated, normal, or delayed aging based on their Age-Gap relative to the model-specific mean absolute error (MAE, rounded up to the nearest whole year): accelerated aging was defined as Age-Gap [≥] +MAE, normal aging as -MAE [≤] Age-Gap < +MAE, and delayed aging as Age-Gap < -MAE. LightGBM provided the best age predictions for both aging clocks, with Pearson correlation coefficients (r) of 0.92 and 0.92 and R-squared values of 0.81 and 0.82 for PPGAge and ABPAge, respectively. Moreover, we found that accelerated aging was significantly associated with higher odds of CVD for both aging clocks, PPGAge (odds ratio, 1.87 [95% CI, 1.59-2.21], p<0.001) and ABPAge (odds ratio, 2.28 [95% CI, 1.51-3.43], p<0.001). Additionally, delayed aging was significantly associated with lower odds of CVD for PPGAge (odds ratio, 0.68 [95% CI, 0.58-0.80], p<0.001). Our findings suggest that aging clocks based on high-fidelity PPG or ABP waveform features may serve as novel biomarkers for identifying individuals with higher odds of CVD.
Related
- Advancing Cardiovascular Nurse Practitioner and Physician Associate Workforce
- Fatty liver index-defined hepatic steatosis and impaired fasting glucose differentially associate with incident diabetes and cardiovascular events in older adults: a nationwide cohort study
- Voice as a biomarker of allostatic load: a systematic review of acoustic features and their association with physiological stress
- Corneal Stiffening and Hyperosmolar Stress Associate With YAP1/TAZ Activation and Ocular Surface Inflammation in Dry Eye Disease.
- Emergent Roles of RVLMC1 Neurons in Cardiovascular Homeostasis and Systemic Physiological Regulation
BibTeX
@misc{pal2026physiological,
title = {{Physiological Waveform Features Based Aging Clocks Associate with Cardiovascular Disease}},
author = {R. Pal and A. Barney and M. Cannesson},
howpublished = {medRxiv},
year = {2026},
month = nov,
doi = {10.64898/2026.11.03.26367198},
url = {https://www.medrxiv.org/content/10.64898/2026.11.03.26367198v1},
note = {CC BY-ND},
}
RIS
TY - UNPB
TI - Physiological Waveform Features Based Aging Clocks Associate with Cardiovascular Disease
AU - R. Pal
AU - A. Barney
AU - M. Cannesson
PY - 2026
DA - 2026/11/03
PB - medRxiv
DO - 10.64898/2026.11.03.26367198
UR - https://www.medrxiv.org/content/10.64898/2026.11.03.26367198v1
AB - Aging affects the morphology of photoplethysmography (PPG) and arterial blood pressure (ABP) waveforms. Here, we demonstrate that machine learning (ML) models trained on features derived from PPG or ABP waveforms can predict an individual's age. We further investigated whether the discrepancy between predicted and chronological age (Age-Gap) could serve as a measure of cardiovascular health by testing its association with cardiovascular disease (CVD). From a large perioperative dataset of 17,327 surgical patients, high-fidelity PPG and ABP waveforms were analyzed, and over 840 time-, frequency-, and statistical-domain features per cardiac cycle were extracted using a signal-processing-based tool. These features were then used to develop ML-based aging clocks, PPGAge and ABPAge, trained on PPG and ABP features, respectively. For each aging clock, we evaluated five ML models: Light Gradient Boosting Machine (LightGBM), eXtreme Gradient Boosting (XGBoost), Random Forest, Lasso, and Elastic Net. Individuals were classified as having accelerated, normal, or delayed aging based on their Age-Gap relative to the model-specific mean absolute error (MAE, rounded up to the nearest whole year): accelerated aging was defined as Age-Gap [≥] +MAE, normal aging as -MAE [≤] Age-Gap < +MAE, and delayed aging as Age-Gap < -MAE. LightGBM provided the best age predictions for both aging clocks, with Pearson correlation coefficients (r) of 0.92 and 0.92 and R-squared values of 0.81 and 0.82 for PPGAge and ABPAge, respectively. Moreover, we found that accelerated aging was significantly associated with higher odds of CVD for both aging clocks, PPGAge (odds ratio, 1.87 [95% CI, 1.59-2.21], p<0.001) and ABPAge (odds ratio, 2.28 [95% CI, 1.51-3.43], p<0.001). Additionally, delayed aging was significantly associated with lower odds of CVD for PPGAge (odds ratio, 0.68 [95% CI, 0.58-0.80], p<0.001). Our findings suggest that aging clocks based on high-fidelity PPG or ABP waveform features may serve as novel biomarkers for identifying individuals with higher odds of CVD.
ER -