Interpretable photoplethysmography-based machine-learning model for noninvasive assessment of systolic blood pressure
Authors:
Open Access
- 1 June 2026
- journal article
- research article
- Published by Frontiers Media SA in Frontiers in Physiology
- Vol. 17, 1779262
Abstract
Background: Hypertension significantly contributes to cardiovascular morbidity and mortality but is often underdiagnosed and poorly controlled, as intermittent cuff-based measurements provide only discrete readings and miss short-term vascular dynamics. Photoplethysmography (PPG), a noninvasive optical technique tracking pulsatile blood-volume changes, offers a practical, low-cost, wearable-compatible alternative capable of capturing temporal and morphological waveform patterns. Unlike dual-sensor or calibration-dependent systems, single-site PPG can extract indices linked to arterial stiffness, compliance, and wave reflection, providing physiologically interpretable vascular information. Advances in signal processing and interpretable machine learning have shifted PPG analysis from descriptive morphology to feature-driven clinical prediction. Integrating waveform metrics—such as systolic amplitude, contour sharpness, and harmonic energy distribution—with demographic variables enables transparent modeling of blood-pressure status. Objective: To develop an interpretable machine-learning model for classifying systolic blood pressure (SBP) status using physiological and morphological features derived from a single PPG signal. Methods: We analyzed 860 participants and extracted waveform-derived and clinical features from 90-second fingertip PPG recordings. A LASSO-regularized logistic regression model with fivefold cross-validation was used for feature selection and classification, and model interpretability was assessed using SHAP values. Results: The final model achieved an AUC of 0.83 (95% CI 0.79–0.87), an F1-score of 0.83, and a Cohen’s d of 1.39. Five predictors were retained: age, BMI, P1 amplitude, waveform sharpness (1_10), and harmonic ratio (H2/H1). Higher P1, greater waveform sharpness, and higher BMI were positively associated with elevated SBP, whereas H2/H1 was negatively associated. Conclusion: An interpretable PPG-based model using physiologically meaningful features can distinguish hypertensive from normotensive individuals and may serve as a practical digital screening tool for elevated systolic blood pressure.Keywords
This publication has 53 references indexed in Scilit:
- Lowering systolic blood pressure to less than 120 mm Hg versus less than 140 mm Hg in patients with high cardiovascular risk with and without diabetes or previous stroke: an open-label, blinded-outcome, randomised trialThe Lancet, 2024
- Genome-wide analysis in over 1 million individuals of European ancestry yields improved polygenic risk scores for blood pressure traitsNature Genetics, 2024
- Blood pressure estimation and classification using a reference signal-less photoplethysmography signal: a deep learning frameworkPhysical and Engineering Sciences in Medicine, 2023
- Photoplethysmograhic sensors, potential and limitations: Is it time for regulation? A comprehensive reviewMeasurement, 2023
- Pharmacological blood pressure lowering for primary and secondary prevention of cardiovascular disease across different levels of blood pressure: an individual participant-level data meta-analysisThe Lancet, 2021
- A multistage deep neural network model for blood pressure estimation using photoplethysmogram signalsComputers in Biology and Medicine, 2020
- Lifetime Risks for Hypertension by Contemporary Guidelines in African American and White Men and WomenJAMA Cardiology, 2019
- Genetic variants in novel pathways influence blood pressure and cardiovascular disease riskNature, 2011
- Global burden of blood-pressure-related disease, 2001The Lancet, 2008
- Effect of increasing heart rate on left ventricular performance in patients with normal cardiac functionThe American Journal of Cardiology, 1988