Reconstructing Cardiovascular Hemodynamics from Non-Invasive Data

35 concepts34 papers2 proposals

To diagnose cardiovascular disease accurately, clinicians traditionally need to look inside the body's transport systems. Whether measuring the pressure drop across a narrowed aorta, mapping the electrical pathways of a misfiring heart, or identifying ischemic tissue, the historical gold standards have relied on invasive procedures. Interventions like cardiac catheterization, contrast angiography, and intracardiac electrophysiology require threading physical probes through the vasculature. These procedures carry inherent risks, including vascular damage, stroke, radiation exposure, and contrast-induced toxicity.

A rapidly growing body of research aims to eliminate this procedural burden. The core intuition driving this field is that the cardiovascular system is a tightly coupled electromechanical and fluidic circuit. A localized pathology—such as a stenotic plaque or a conduction block—cannot occur in isolation; it leaves subtle, deterministic footprints on the system's macroscopic behavior. By treating non-invasive observations (like surface ECGs, ultrasound videos, and non-contrast imaging) as inputs to an inverse problem, modern computational models can reconstruct internal hemodynamic and physiological states without ever penetrating the skin.


The Core Intuition: Solving the Inverse Problem

Why can a machine learning model or a fluid simulation deduce internal pressures or tissue states from the outside? Consider blood flowing through an artery. A physical narrowing (stenosis) alters the velocity profile, wall shear stress, and downstream turbulence of the blood. These fluid dynamics directly influence how the heart muscle contracts and vibrates, which can be captured via ultrasound or magnetic resonance imaging.

Similarly, an electrical block in the heart's conduction system alters the precise sequence of myocardial contraction, which subtly changes the shape of a standard electrocardiogram (ECG) or the frame-by-frame deformation of an echocardiogram video.

Instead of physically inserting a sensor to measure these phenomena, researchers use deep learning and physics-informed algorithms to decode these surface signatures. This transforms non-invasive data from a simple visualization tool into a rich, quantitative diagnostic medium.


Decoding Hemodynamics via Fluid Dynamics and Machine Learning

One of the most direct applications of this approach is estimating internal blood pressure gradients and fractional flow reserves—metrics that traditionally require a pressure-wire catheter.

To bypass catheterization in conditions like aortic coarctation (a narrowing of the large blood vessel branching from the heart), Kristóf Ralovich et al. demonstrated a pipeline that combines machine learning-based MRI segmentation with Computational Fluid Dynamics (CFD). By extracting patient-specific anatomical boundaries from non-invasive MRIs and applying numerical fluid simulations, the authors successfully estimated internal blood pressure gradients, matching the accuracy of invasive catheterization.

[Non-Invasive Imaging (MRI/CT)] ──> [ML Segmentation] ──> [CFD / Physics Models] ──> [Internal Pressure Gradients]

While traditional CFD simulations are highly accurate, they are computationally expensive and time-consuming. To bridge this gap for real-time clinical triage, Young‐Woo Kim et al. developed a two-step machine learning framework. By training models on synthetic data and patient biometrics, they bypassed the heavy computational overhead of raw fluid simulations, replicating invasive pressure measurements rapidly and non-invasively to accelerate clinical decision-making.


Extracting Latent Signals from Motion and Radiomics

When fluid dynamics or structural changes are too subtle for the human eye to detect on a standard scan, deep learning models can extract "radiomic" features—complex spatial and temporal patterns across pixels or voxels.

For example, identifying a myocardial infarction (heart attack) typically requires gadolinium-based contrast agents, which carry risks of neurotoxicity and nephrotoxicity. To eliminate this risk, Chenchu Xu et al. developed a deep learning framework called OF-RNN. Instead of looking for contrast enhancement, the model analyzes microscopic, pixel-level motion abnormalities in standard, contrast-free cardiac MRI sequences, achieving a 94.35% accuracy in locating infarcted tissue purely from mechanical motion signatures.

This strategy of converting video motion into internal pressure metrics is also vital for fragile populations, such as newborns. Diagnosing neonatal pulmonary hypertension usually demands invasive heart catheterization. Addressing this, Hanna Ragnarsdóttir et al. introduced a multi-view 3D convolutional neural network that processes standard echocardiogram videos. The model automatically learns the spatio-temporal patterns of the infant's heart wall motion to detect and classify the severity of pulmonary hypertension non-invasively.

Similarly, severe coronary artery stenosis can lurk behind an apparently normal ECG during routine checkups. Zheng-Kai Xue et al. trained a deep transfer learning model on 12-lead ECGs combined with basic clinical data. The network detects latent electrophysiological anomalies caused by hidden arterial narrowing, providing a cheap, non-invasive early warning system before a catastrophic cardiac event occurs.


Mapping Electrical and Metabolic Pathways Remotely

The frontier of non-invasive reconstruction extends beyond fluid pressures and tissue structure into tracking the heart's internal electrical conduction and metabolic shifts.

For cardiac resynchronization therapy to work, clinicians must locate precise electrical conduction failures. Rather than using invasive electroanatomical mapping catheters, Adityo Prakosa proposed an inverse electro-kinematic learning framework. By leveraging synthetic 3D echocardiographic sequences where both electrical activation and mechanical motion are known, the author trained a machine learning model to estimate internal cardiac electrical propagation patterns directly from non-invasive ultrasound images.

[Surface Observations (ECG / Ultrasound)] ──> [Inverse Kinematic Model] ──> [Internal Electrical Propagation Map]

On the micro-signal scale, monitoring the heart's internal conduction system during valve replacements typically requires intracardiac catheters. Nolwenn Tan et al. overcame this by utilizing high-resolution body surface potential mappings. By combining specialized signal averaging during the exhalation phase with localized electrode configurations, the authors successfully isolated the "His bundle potential"—a minute internal electrical signal—from the surface of the skin, matching invasive gold standards.

Even metabolic shifts can be tracked safely. In hemodialysis patients, traditional stress tests for coronary artery disease are highly dangerous. Masato Nishimura et al. demonstrated that $^{123}$I-BMIPP myocardial scintigraphy—a resting imaging technique—can accurately detect coronary artery disease by tracking shifts in fatty acid metabolism rather than forcing a high-risk patient through physical or pharmacological stress.


Conclusion

The shift toward non-invasive inference represents a fundamental change in cardiovascular medicine. As highlighted in a consensus statement by Marc Dewey et al., while clinical guidelines strongly advocate for quantitative physiological assessments, actual adoption of invasive measurements remains low globally due to risk, cost, and complexity.

By treating the cardiovascular system as a predictable physical network, these diverse computational approaches prove that the internal state of the heart can be accurately mapped from the outside. Whether through fluid simulations, radiomic feature extraction, or high-resolution surface signal processing, the integration of these non-invasive tools promises to make high-fidelity cardiac diagnostics safer, faster, and universally accessible.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_invasive_diagnostic_risk_552a11ff,
  title = {invasive diagnostic risk},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/invasive-diagnostic-risk/rdr_552a11ff521640d7a15c2778ed8689d5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_invasive_diagnostic_risk_552a11ff_v1,
  title = {invasive diagnostic risk},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/invasive-diagnostic-risk/rdr_552a11ff521640d7a15c2778ed8689d5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake