Abstract: Photoplethysmography (PPG) has emerged as one of the most widely adopted modalities for wearable vital-sign monitoring due to its low cost, simplicity, and compatibility with portable and ambulatory devices. However, extracting accurate and physiologically meaningful vascular information from PPG signals remains challenging because of their nonlinear, subject-dependent nature and susceptibility to noise and motion artifacts. This work presents a series of data-driven approaches for continuous vascular health assessment using PPG signals. Initially, machine learning models were investigated for blood pressure classification into hypotension, normotension, and hypertension categories, achieving accuracies up to 87%. Due to the requirement of manually identifying fiducial points from PPG cycles, these experiments were conducted on a curated subset of approximately 1,627 subjects from the PulseDB dataset. Building on this, deep learning approaches were developed for continuous blood pressure estimation, where systematic analysis showed that incorporating the first and second derivatives of PPG improves performance, achieving mean absolute errors (MAEs) as low as 4.11 mmHg for systolic blood pressure (SBP) and 3.53 mmHg for diastolic blood pressure (DBP). A lightweight Modified Half U-Net architecture was proposed, achieving competitive performance with reduced computational complexity, with MAEs of 4.15 mmHg (SBP) and 3.71 mmHg (DBP). Beyond point-wise blood pressure estimation, a hybrid Transformer–CNN framework was developed for continuous ABP waveform reconstruction, achieving reconstruction errors of 4.56 ± 5.72 mmHg along with SBP and DBP estimation errors of 4.79 mmHg and 3.93 mmHg, respectively. The blood pressure estimation and arterial blood pressure (ABP) waveform reconstruction models were developed and evaluated on the PulseDB dataset comprising 4,091 filtered subjects obtained after physiological quality screening from the original 5,361-subject cohort.

Arterial stiffness assessment was further investigated through pulse wave velocity (PWV) estimation using a CNN–Transformer model, achieving an MAE of 0.68 m/s, RMSE of 1.20 m/s, and R² of 0.80. Finally, a Unified Vascular Score (UVS) was proposed to integrate multiple vascular health markers into a single interpretable representation of vascular state, demonstrating strong agreement with conventional cardiovascular risk frameworks while additionally revealing vascular heterogeneity within the same risk category. The UVS framework was developed using data from 1,303 subjects and further compared against traditional risk scores using an independent cohort of 205 participants. While the proposed methods demonstrated promising performance across multiple vascular assessment tasks, further validation on diverse ambulatory and community-based populations is necessary to establish robustness in real-world wearable settings. Additionally, the PWV estimates used in this work are surrogate measures rather than gold-standard carotid–femoral PWV measurements. Together, these contributions support scalable and wearable approaches for continuous vascular health monitoring and preventive cardiovascular assessment.

Keywords: Photoplethysmography, Blood Pressure Estimation, Arterial Blood Pressure Reconstruction, Pulse Wave Velocity, Arterial Stiffness, Deep Learning, Unified Vascular Score

All are cordially invited.

Event Details
Title: Data-Driven Approaches for Assessment of Vascular Health Markers: Potential for Ambulatory Monitoring
Date: May 21, 2026 at 03:00 PM
Venue: ESB 244
Speaker: Mr. Mathew Cigi (EE23S031)
Guide: Dr. Arunkumar D Mahindrakar
Type: MS seminar

Updated: