Smart Controller Turns Cardiovascular Models Into Reliable Digital Twins for Childhood Sepsis

Sepsis remains one of the most feared adversaries in paediatric intensive care, accounting for more than eight percent of admissions to paediatric intensive care units and contributing heavily to childhood mortality worldwide. Yet even with modern monitors tracking blood pressures, heart rate and cardiac output, clinicians are often flying blind: the underlying causes of a child’s circulatory collapse, such as how much blood volume has been lost from the vessels, how stiff the failing heart has become, or how far the vasculature has dilated, cannot be measured directly at the bedside. A new study published in PLOS Computational Biology tackles this hidden-variable problem head-on, introducing a calibration framework that lets mechanistic cardiovascular models lock onto a single, physiologically correct solution from limited clinical data.

The work, led by Manuel T. Cabeleira of University College London together with Samiran Ray of Great Ormond Street Hospital, Nicholas C. Ovenden and Vanessa Díaz-Zuccarini, addresses a long-standing bottleneck in computational physiology. Lumped-parameter cardiovascular models, which represent the circulation as a network of compartments connected by resistances and compliances governed by ordinary differential equations, can mimic physiological behaviour with impressive precision. But calibrating them, meaning finding the parameter values that reproduce a specific patient’s haemodynamics, has proven notoriously difficult. Because a single observable such as arterial pressure is influenced simultaneously by many parameters, traditional optimisation methods typically return a multitude of different parameter sets that all satisfy the same targets, or they require enormous computational searches that scale poorly with model size.

The team’s solution is elegant in conception: the Embedded Feedback Controller, or EFC. Instead of treating model parameters as fixed constants to be adjusted by an external optimiser, EFC promotes selected physiological parameters into dynamic states with their own differential equations. Each controller equation exploits the qualitative structure of the governing physiology, the monotonic correlations between parameters and variables that can be read directly from the model equations, to nudge its assigned parameter in the direction that reduces the error against a prescribed clinical target. Compliance, resistance and elastance parameters essentially steer themselves, continuously, until the entire circulation settles into a state where pressures, flows and volumes all match the prescribed values.

This embedded approach solves the ambiguity problem through over-determination. Sixteen parameters are calibrated against sixteen targets, but those targets are not sixteen independent constraints. Eight of them are compartmental volumes that must sum to a fixed total blood volume, and single pressure waveforms carry multiple conditions at once. Because every parameter influences several compartments, each controller imposes a global constraint on the circulation, and the only configuration satisfying all of them simultaneously is the one at which every relative error vanishes. The result is a single, unique solution, reached reproducibly from any starting point. In a stress test of 1024 simulations with randomly initialised parameters and volume distributions, every single run converged to the same answer, with parameter variability across runs showing a median coefficient of variation of just 0.016 percent.

The benchmarks against established methods are striking. Gradient descent, with 64 multi-starts, achieved the lowest best-case error at 0.04 percent, but 35 of its 64 restarts stalled in a spurious local minimum at roughly half the correct systemic arterial resistance, missing targets by up to 18 percent, with nothing in the optimiser’s own output revealing the failure. Markov chain Monte Carlo, the gold standard for Bayesian inference, required roughly 192,000 forward model solves and delivered a posterior spread 70 times wider per target than the linear EFC. The linear EFC, by contrast, needed only 370 solves per calibration, finishing a single virtual patient in under fifty seconds on a workstation CPU, and placed every one of its 1024 runs within one percent of every target. Notably, the MCMC posterior concentrated on exactly the region the controller calibrations found, independent confirmation that EFC converges on the true solution.

To demonstrate clinical relevance, the researchers applied the framework to paediatric septic shock, a particularly demanding test case because the disease manifests as two distinct haemodynamic phenotypes. Warm shock is characterised by massive peripheral vasodilation with preserved or elevated cardiac output, while cold shock involves vasoconstriction and reduced cardiac output; the two can produce confusingly similar bedside pictures despite requiring different treatments. Using literature-derived target ranges for a reference twelve-month-old child, the team generated three virtual populations of 200 subjects each: normal physiology, warm shock and cold shock. Across all 600 virtual subjects, residual errors stayed below one percent for pressures, flows and compartmental volumes.

The resulting parameter distributions reproduce the expected haemodynamic fingerprints of each phenotype. The warm shock cohort occupied consistently low systemic resistance regimes, consistent with near-maximal vasodilation, and showed higher total compliance than cold shock. The cold shock population spanned higher and more variable resistance states, and at normal blood volumes showed markedly reduced systolic elastance with preserved diastolic elastance, a signature compatible with impaired contractile strength and a cardiogenic component to the shock. Because every virtual subject is calibrated through the same physiological structure, differences between populations can be attributed directly to underlying adaptations in vascular resistance, compliance, cardiac elastance and effective blood volume, rather than to artefacts of the calibration method itself.

Perhaps the most conceptually important finding is what happens when the framework fails. When a prescribed combination of targets is structurally incompatible with the model, the corresponding controller drives its parameter to the edge of its permissible range and comes to rest there, still carrying a residual error. Rather than masking the problem, the calibration makes the infeasibility explicit: a pinned parameter with a persistent residual is a readable diagnostic that the requested physiology lies outside what the model can produce. In the virtual septic populations, such saturation appeared in specific regimes, for instance on the unstressed pulmonary arterial volume in warm shock, while the majority of subjects converged cleanly. This reframing of residual error, from optimisation failure to feasibility diagnostic, aligns directly with recent methodological guidelines calling for mechanistic models that expose identifiability limits rather than hide them behind black-box optimisation.

The authors are candid about limitations. The study is entirely computational, built from literature-derived targets rather than patient recordings, and individual virtual subjects may combine values that are plausible in isolation yet rarely co-occur in a single child. Calibration currently works on steady-state snapshots using trend-level monitor data, the model omits blood-flow inertia and any explicit autonomic baroreflex, and the septic physiology represented is confined to haemodynamics, without the metabolic derangements such as lactate elevation that also define the disease. Even so, the path forward is built into the method. Because each calibration provides a natural initialisation for the next, the framework extends naturally to longitudinal data, tracking a patient’s evolving physiology across an ICU stay. That capability, combined with the speed, uniqueness and interpretability demonstrated here, positions EFC as a practical foundation for the long-promised goal of patient-specific cardiovascular digital twins in critical care, where a model calibrated in minutes could help clinicians infer the invisible mechanisms driving a child’s shock and test treatment strategies before touching the patient.

Subject of Research: A physiologically constrained calibration framework for lumped-parameter cardiovascular models applied to generating virtual paediatric sepsis populations

Article Title: A physiologically constrained calibration framework for cardiovascular models applied to a paediatric synthetic sepsis population

Article References: Cabeleira, M. T., Ray, S., Ovenden, N. C., & Díaz-Zuccarini, V. (2026). A physiologically constrained calibration framework for cardiovascular models applied to a paediatric synthetic sepsis population. PLOS Computational Biology, 22(10), e1014847. https://doi.org/10.1371/journal.pcbi.1014847

Image Credits: AI Generated

DOI: 10.1371/journal.pcbi.1014847

Keywords: cardiovascular modelling, sepsis, paediatric intensive care, digital twins, model calibration, septic shock, lumped-parameter models, virtual populations, computational physiology, haemodynamics, ordinary differential equations, critical care