Hybrid Modeling of Photoplethysmography for Non-invasive Monitoring of Cardiovascular Parameters

TL;DR

This work tackles non-invasive estimation of cardiac biomarkers, notably stroke volume and cardiac output, from finger photoplethysmography (PPG) by bridging simulations and real data in a probabilistic framework. It introduces a hybrid approach that uses a conditional variational auto-encoder to generate plausible arterial pressure waves (APWs) from PPG and a neural posterior estimator trained on simulated APW data to infer biomarkers, combining them via $p(\theta \mid \mathbf{y}) \approx \mathbb{E}_{p(\mathbf{x} \mid \mathbf{y})}[p(\theta \mid \mathbf{x})]$. On unseen VitalDB data, the method tracks temporal changes in SV and CO with higher per-subject correlations than baselines, and benefits from the generative PPG-to-APW mapping by capturing ambiguity and uncertainty. Absolute-value accuracy remains challenging, underscoring the need for personalization and calibration; nonetheless, the framework demonstrates a practical path toward non-invasive, physics-informed cardiac monitoring suitable for wearables and clinical use.

Abstract

Continuous cardiovascular monitoring can play a key role in precision health. However, some fundamental cardiac biomarkers of interest, including stroke volume and cardiac output, require invasive measurements, e.g., arterial pressure waveforms (APW). As a non-invasive alternative, photoplethysmography (PPG) measurements are routinely collected in hospital settings. Unfortunately, the prediction of key cardiac biomarkers from PPG instead of APW remains an open challenge, further complicated by the scarcity of annotated PPG measurements. As a solution, we propose a hybrid approach that uses hemodynamic simulations and unlabeled clinical data to estimate cardiovascular biomarkers directly from PPG signals. Our hybrid model combines a conditional variational autoencoder trained on paired PPG-APW data with a conditional density estimator of cardiac biomarkers trained on labeled simulated APW segments. As a key result, our experiments demonstrate that the proposed approach can detect fluctuations of cardiac output and stroke volume and outperform a supervised baseline in monitoring temporal changes in these biomarkers.

Figures (4)

• Figure 1: Proposed hybrid framework for non-invasive monitoring of cardiovascular parameters from PPG signals. The framework combines simulated and real-world data across two key components: a) a conditional VAE, trained on real-world paired APW and PPG segments, producing plausible APW signals from PPG inputs; b) a density estimator trained on a in-silico dataset of hemodynamic simulations inferring cardiovascular biomarkers from generated APW signals. The framework is validated in-vivo on unseen VitalDB data, using per-subject Spearman correlation between ground truth and predicted values of cardiovascular biomarkers.
• Figure 2: Box plots illustrating the per-subject Spearman correlation coefficients for the estimation of four key cardiovascular parameters. Our proposed hybrid approach (green) is compared against three relevant baselines: (i) APW (orange), which uses the ground-truth invasive APWs for inference with estimator trained in-silico; (ii) PPG Windkessel (pink), an estimator trained on labeled in-silico PPGs obtained with a Windkessel model approximation Manduchi2024Leveraging; and (iii) PPG Supervised (blue), an estimator trained directly on a subset of labeled VitalDB PPG data. Average results across independent runs with standard deviations are reported in \ref{['tab:ext-num-results']}.
• Figure 3: SV ground truth (gray), predicted values with our approach (green), and redictions obtained with invasive inference from ground truth APW (orange), on different subjects from the VitalDB dataset.
• Figure 4: Rejecting unusable samples via predicted uncertainty for HR inference. (a) Twelve VitalDB PPG segments with the highest predicted uncertainty. (b) Twelve VitalDB PPG segments with the least predicted uncertainty. All PPG signals underwent bandpass filtering.