Building a Mouse Cardiopulmonary Digital Twin from Sparse Data

In a development that could reshape how scientists study heart-lung disease, a team of researchers has built a working digital twin of the mouse cardiopulmonary system, a computational replica capable of extracting detailed information about blood flow and heart function from remarkably sparse measurements. The study, published in Annals of Biomedical Engineering, demonstrates that a relatively simple mathematical model, fitted to pressure and volume data from individual animals, can infer quantities that would otherwise be unmeasurable in a living mouse. The work, led by Vitaly O. Kheyfets of the University of Colorado Anschutz Medical Campus together with Kenzo Ichimura and Edda Spiekerkoetter of Stanford University, along with colleagues at Yale School of Medicine, lays the groundwork for a new generation of “in silico” experiments in which disease progression and potential treatments are tested first inside a computer rather than inside an animal.

Rodent models are the backbone of cardiopulmonary research. From pulmonary hypertension to right ventricular failure, most of what is known about how the heart and lung vasculature adapt under disease has been learned by studying mice and rats. Yet the very smallness that makes mice convenient also makes them extraordinarily difficult to instrument. A mouse heart weighs less than a paperclip and beats more than five hundred times per minute, and comprehensive functional characterization of each animal’s circulatory system is often impractical. Researchers typically must choose between a handful of partial measurements, and integrating multimodal data across experiments remains a persistent challenge. The new study confronts this problem directly by asking how much physiological insight can be squeezed from the data scientists already collect.

The team’s approach centers on a so-called zero-dimensional, or 0D, model, a lumped-parameter framework that represents the cardiopulmonary axis not as a detailed three-dimensional anatomy but as a network of resistances, compliances, and pressure-volume relationships bounded by the right and left atria. In such models, the vasculature is abstracted into elements that describe how blood vessels oppose and store flow, while the ventricle is described by time-varying elastance curves that capture its contractile behavior through the cardiac cycle. The researchers constructed a model incorporating thirteen unknown parameters representing pulmonary vascular impedance, right ventricular pressure-volume dynamics, and the behavior of the tricuspid and pulmonic valves, including regurgitation, the backward leakage of blood through imperfectly closing valves that is common in diseased hearts.

A crucial step in the work was determining which of those thirteen parameters could actually be estimated from the available data, a question of what mathematicians call identifiability. Through sensitivity and identifiability analyses, the team discovered that a reduced subset of nine parameters was both structurally and practically identifiable, meaning that the pressure and volume waveforms contained enough information to constrain those nine values uniquely, while the remaining four could not be reliably inferred and were excluded from the fitting procedure. This distinction matters because attempts to optimize unidentifiable parameters can produce models that fit the data but generate physically meaningless parameter values, undermining confidence in every other quantity the model reports. By winnowing the parameter set before optimization, the researchers ensured that their digital twin produced robust, interpretable results.

The model was tested against data from twenty-eight mice spanning four distinct surgical conditions, including two scenarios of mechanically induced right ventricular pressure overload, conditions that mimic the hemodynamic burden seen in pulmonary hypertension. When the identifiable parameters were optimized to fit each animal’s measurements, the digital twin reproduced the observed physiology with striking fidelity. The model’s predictions of the maximum rate of pressure rise in the right ventricle, a key index of contractility known as dP/dt, correlated with the measured values at r = 0.94, with limits of agreement tighter than one millimeter of mercury per second. Predictions of ventricular volume extremes were even better: end-diastolic volume showed a correlation of r = 0.97 and end-systolic volume reached r = 0.98, with limits of agreement of roughly five microliters for both, a margin comparable to the resolution of the underlying conductance catheter measurements in this small animal setting.

Beyond merely fitting the data, the model extracted physiologically meaningful quantities that no single measurement could provide directly. As expected, mice subjected to right ventricular pressure overload exhibited elevated inferred values of end-systolic elastance, end-diastolic elastance, and right atrial pressure, reflecting both the increased contractile demands placed on the overloaded ventricle and the rising filling pressures that accompany disease. Notably, the model-inferred end-systolic elastance, considered a load-independent measure of contractility, correlated moderately with conventional single-beat estimates of right ventricular contractility at r = 0.68, a statistically significant relationship that suggests the digital twin is capturing genuine contractile physiology rather than mathematical artifacts. One animal whose inferred parameters fell outside physiological plausibility was excluded from the analysis, an example of how the framework can also flag suspect data.

The concept of a digital twin, a continuously updated computational replica of a specific biological system, has gained traction in human medicine, where patient-specific heart models are being explored for surgical planning and device testing. Translating that vision to the laboratory mouse, however, demands a different philosophy. Where clinical models can draw on imaging, catheterization, and longitudinal follow-up, mouse studies often yield only sparse, terminal measurements. The Kheyfets team’s contribution is to show that a 0D model, deliberately kept simple enough to be constrained by such limited data, can still deliver individualized hemodynamic insight. The right ventricular pressure waveform, long treated as a secondary signal in animal studies, emerges here as a rich source of information about the coupled heart-lung system.

The implications for drug development and disease modeling are considerable. Pulmonary hypertension research in particular depends on quantifying how the right ventricle adapts, or fails to adapt, to increased afterload, and how experimental therapies improve ventricular-vascular coupling. Traditional indices such as pulmonary vascular resistance capture only part of this picture; the interplay between vascular impedance, ventricular elastance, and valvular function determines outcomes in ways that single metrics obscure. A validated digital twin allows researchers to run counterfactual simulations on the same animal, asking, for instance, how a given reduction in vascular stiffness would translate into changes in ventricular workload, all without touching the animal. This capacity for in silico testing of interventions could reduce animal numbers, accelerate preclinical screening, and illuminate mechanisms that are invisible to direct measurement.

The study also contributes methodologically to the broader field of physiological modeling. Identifiability analysis, profile-likelihood approaches, and modern sensitivity techniques, including Shapley-value-based attribution methods adapted from machine learning, are increasingly being used to make biological models trustworthy. By formally demonstrating which parameters of a cardiopulmonary 0D model can and cannot be inferred from right ventricular pressure-volume data, the authors provide a practical template for other groups building subject-specific models in small animals. The framework’s reliance on standard hemodynamic measurements means it could be adopted by laboratories without specialized computational infrastructure, and the authors state that all data and code from the project will be made publicly available on GitHub, lowering the barrier to adoption and replication.

The authors caution that the approach is a foundation rather than a finished clinical tool. The model was validated within a specific experimental context, and its inferences, while physiologically plausible and externally consistent with established measures of contractility, ultimately depend on the assumptions embedded in the lumped-parameter structure. Extending the framework to other disease models, integrating additional data streams such as imaging, and determining how well inferred parameters track disease over time remain open questions. Still, the central result stands: subject-specific computational modeling enables inference of ventricular function and pulmonary hemodynamics from sparse right ventricular pressure and volume data, providing access to otherwise unmeasurable quantities. For a field long constrained by what can be physically measured in a creature the size of a thumb, the digital twin offers a way to see the invisible, and it points toward a future where every experimental animal carries, in parallel with its biological existence, a mathematical counterpart capable of answering questions the laboratory alone cannot.