Leveraging Cardiovascular Simulations for In-Vivo Prediction of Cardiac Biomarkers

TL;DR

The paper tackles the inverse problem of inferring cardiovascular biomarkers from biosignals by pairing a whole-body 1D cardiovascular simulator with simulation-based inference (SBI) using amortized neural posterior estimation. It introduces a stochastic measurement model to bridge sim-to-real gaps, trains a neural density estimator on a large synthetic dataset, and validates the approach both in silico and in vivo on VitalDB data, providing per-sample uncertainty for biomarkers such as HR, CO, SVR, and LVET. A hybrid learning strategy blends synthetic data with a small real calibration set to mitigate model misspecification and improve in-vivo performance. The framework enables population-level uncertainty analyses, robust temporal tracking of CO/SVR, and automatic rejection of low-quality measurements, offering a foundation for uncertainty-aware, non-invasive cardiovascular monitoring and personalized medicine.

Abstract

Whole-body hemodynamics simulators, which model blood flow and pressure waveforms as functions of physiological parameters, are now essential tools for studying cardiovascular systems. However, solving the corresponding inverse problem of mapping observations (e.g., arterial pressure waveforms at specific locations in the arterial network) back to plausible physiological parameters remains challenging. Leveraging recent advances in simulation-based inference, we cast this problem as statistical inference by training an amortized neural posterior estimator on a newly built large dataset of cardiac simulations that we publicly release. To better align simulated data with real-world measurements, we incorporate stochastic elements modeling exogenous effects. The proposed framework can further integrate in-vivo data sources to refine its predictive capabilities on real-world data. In silico, we demonstrate that the proposed framework enables finely quantifying uncertainty associated with individual measurements, allowing trustworthy prediction of four biomarkers of clinical interest--namely Heart Rate, Cardiac Output, Systemic Vascular Resistance, and Left Ventricular Ejection Time--from arterial pressure waveforms and photoplethysmograms. Furthermore, we validate the framework in vivo, where our method accurately captures temporal trends in CO and SVR monitoring on the VitalDB dataset. Finally, the predictive error made by the model monotonically increases with the predicted uncertainty, thereby directly supporting the automatic rejection of unusable measurements.

Figures (17)

• Figure 1: Proposed framework for predicting cardiovascular parameters from cardiac measurements (biosignals). Results are shown using arterial radial pressure waveforms (PWs), though PPGs can also be used. A: We generate a large-scale dataset of CV simulations using whole-body 1D hemodynamics simulators and appropriate noise models (see \ref{['fig:pre-processing']} for more details). B: A neural posterior estimator is trained on the in-silico dataset. It learns a surrogate of the posterior distribution of the parameters of interest given the PWs. C: The in-vivo dataset is comprised of PWs used to test the model on real-world distributions (gray histogram)and a small calibration set with labels used to finetune part of the model (pink histogram). D: The proposed framework is tested in-silico. The corner plot shows the learned posterior distributions of cardiac biomarkers for three colored waveforms from panel A, the prior distribution is shown in gray. E: The proposed framework is further validated in-vivo. The upper plot shows results on the model trained in-silico, while the lower plot shows results on the model further fine-tuned on the calibration set. For each plot, we show (left) the histogram of per-patient Spearman correlation between the measured and predicted CO, and (right) the normalized predicted vs measured CO of a patient through time.
• Figure 2: MAE and ACAUC of the learned posterior distribution at different SNR. Results are reported for the NPE algorithm trained on synthetic APWs with (a) no noise model, (b) stochastic noise, and (c) fixed high noise.
• Figure 3: Average SCI, for credibility levels $68\%$ and $95\%$. The x-axis denotes the SNR for both APWs and PPGs.
• Figure 4: Left: histogram of the per-patient Spearman correlations between the measured CO in VitalDB and the predicted CO by NPE, which is trained solely on simulated APWs. Right: we plot the normalized CO tracking provided by VitalDB (True CO) together with the NPE prediction (Pred CO) for patients with different correlations.
• Figure 5: Uncertainty analysis of the NPE trained on synthetic APWs and tested on VitalDB APWs. Left: Three VitalDB APWs characterized by different cardiac functions and noise in the recording. Right: Corner plot showing the learned posterior distributions of the corresponding APWs and their prior distribution (grey).