Heart muscle is a woven material: sheets of fibers that stretch, slide and shear past one another while the tissue keeps its volume almost exactly constant. Computer models of the heart describe this with a strain energy, a formula that says how much energy each kind of deformation stores. They then fit the formula's stiffness constants to patients' MRI scans. A careful sensitivity study of one such model found that only the bulk and along-fiber constants matter for the quantities MRI measures. The cross-fiber and fiber-sheet shear constants seemed to have no effect and could be dropped. We think this conclusion may be partly an artifact. Two things could hide shear. The first is volumetric locking, a numerical failure in which ordinary finite elements become artificially stiff when the material is nearly incompressible and can no longer deform freely. The second is the fact that cardiac models treat fiber tension and shear as independent. Textile engineers learned long ago that they are not: pulling on a fiber raises the shear stiffness around it.
This note proposes a controlled test of that suspicion, followed by one narrow use of the answer. The test repeats the sensitivity analysis while switching the numerical formulation between a locking-prone one and a locking-free one, and switching a fabric-style tension-shear coupling term on and off. If shear parameters become identifiable under the locking-free, coupled model, we then ask a practical genetics question. Titin (TTN) variants of uncertain clinical significance are a large source of unresolved genetic diagnoses, and titin is the protein spring that bears tension in heart muscle. We ask whether a stiffness parameter that does not depend on how hard the heart is loaded separates titin variant carriers from non-carriers better than the raw MRI strain already used in population genetics. The core claim is falsifiable at both stages. If either stage fails, we report the failure, because a negative answer is still useful to cardiac modelers and to geneticists.
Core Claim and Motivation
The study that motivates this work is Sensitivity analysis and inverse uncertainty quantification for the left ventricular passive mechanics. It trained Gaussian process emulators on 2,000 finite element runs of one healthy left ventricle during passive filling. An emulator here is a cheap statistical stand-in for the full simulation, and these reached $Q^2$ of 0.98–1.00. The study then found, using Sobol indices (a variance-based measure of how much each input drives each output), that only four of the eight Holzapfel-Ogden constants affect cavity volume and segmental strains. The shear and sheet constants $a_s, b_s, a_{fs}, b_{fs}$ were negligible. That study does not examine the discretization, and its loading is low-pressure passive filling.
The textile forming literature points the other way. The review Material Modelling of Fabric Deformation in Forming Simulation of Fiber-Metal Laminates and Modeling multiaxial stress states in forming simulation of woven fabrics show that dropping tension-shear coupling systematically underestimates shear angles under constrained loading. Constrained loading means the fabric cannot escape by wrinkling, which is roughly the situation of myocardium confined in a closed, nearly incompressible wall. The mechanism they describe is concrete: tension straightens a fiber, which raises contact force and friction where it crosses neighbors, so shear stiffness becomes a function of fiber tension. The work on non-crimp fabrics, Investigation of the Membrane Behavior of UD-NCF in Macroscopic Forming Simulations, likewise finds that coupling transverse tension to shear changes predicted defects. We borrow this mechanism. We do not claim heart tissue has yarn crossovers. We claim only that a fiber-tension-dependent shear stiffness is a plausible phenomenological form worth testing against the decoupled one.
The numerical side is supplied by An accurate, robust, and efficient finite element framework with applications to anisotropic, nearly and fully incompressible elasticity. That paper shows three things in artery and biventricular benchmarks. Penalty formulations lock. The standard split of the deformation gradient into volumetric and shape-changing parts produces spurious fiber stresses in anisotropic tissue. Pressure-projection-stabilized equal-order elements and MINI elements fix both problems. It used nominal literature parameters and never asked whether locking changes which parameters can be inferred from data. That question is the one we add.
Concretely, we augment the fiber-sheet shear term of the Holzapfel-Ogden energy with a tension-dependent prefactor:
$$ \Psi_{fs} = \frac{a_{fs}\,\big(1 + \kappa \,\langle I_{4f} - 1 \rangle_+\big)}{2 b_{fs}} \left[\exp\!\big(b_{fs} I_{8fs}^2\big) - 1\right]. $$
Here $I_{4f}$ is the squared fiber stretch and $I_{8fs}$ is the fiber-sheet shear invariant. The constants $a_{fs}, b_{fs}$ are the usual shear stiffness constants. $\langle\cdot\rangle_+$ keeps only positive values, so the term switches on only when the fiber is in tension. $\kappa \ge 0$ is the new coupling strength. Setting $\kappa = 0$ recovers the standard model exactly, which makes the coupling an ablatable ingredient. Before any cardiac use, we check that this term behaves sensibly. We fit $\kappa$ against the qualitative off-axis and bias-tension behavior reported in Membrane behavior of uni- and bidirectional non-crimp fabrics in off-axis-tension tests and Origin of tensile strength of a woven sample cut in bias directions. This is a code-verification step, not a claim about fabrics.
Research Program
The first stage is a factorial identifiability study. We cross two discretizations with two energies. The discretizations are penalty displacement elements and the stabilized P1-P1 formulation of Karabelas et al., with MINI elements as a secondary check. The energies are decoupled Holzapfel-Ogden ($\kappa=0$) and coupled ($\kappa$ free). We run these on the patient-specific passive filling geometry used by the sensitivity study. We also run them under a second loading regime: end-systole with an active fiber stress. Active contraction loads fibers in tension, so it is the regime where coupling should matter. For each of the four cells, we train emulators on a matched design and recompute Sobol indices. We compute them not only for cavity volume and segmental circumferential strain, but also for longitudinal strain, circumferential-longitudinal shear strain and torsion, since shear-type observables are where shear parameters would show up. The quantity of interest is the change in total-order Sobol index for $a_{fs}, b_{fs}, \kappa$ across cells.
This design separates the competing explanations. If the indices rise under the locking-free formulation even with $\kappa=0$, locking was hiding shear. If they rise only with $\kappa>0$ and active loading, the cause is a missing coupling term excited by tension. If they stay negligible everywhere, the original conclusion stands and shear parameters really are unidentifiable from clinical strain.
The second stage, conditional on the first, does inverse inference on real subjects. We use UK Biobank cardiac MRI with exome data, as analyzed in Mapping rare protein-coding variants on multi-organ imaging traits. That study found burden associations between rare TTN variants and regional peak circumferential strain, using AlphaMissense to group damaging missense variants. We select carriers of predicted-damaging TTN variants and matched non-carriers, matched on age, sex, body size and cuff blood pressure. We segment with automated quality control following Automated Quality Control in Image Segmentation: Application to the UK Biobank Cardiac MR Imaging Study. We then fit the identifiable parameters per subject with emulator-accelerated Bayesian inversion. Per-subject full finite element solves are too costly, so emulators are trained over geometry shape modes plus material parameters. The Bayesian inversion uses the No-U-Turn Sampler approach of the sensitivity study, and derivative-free strain matching as in Strain-Based Regional Nonlinear Cardiac Material Properties Estimation from Medical Images serves as a cross-check.
The hypothesis that justifies adding mechanics here is narrow. Strain mixes material stiffness with loading and geometry. A fitted material parameter should, in principle, remove the loading and geometry contributions. So it should separate carriers from non-carriers more cleanly, especially in subjects whose loading differs. We treat ventricular pressure as an uncertain input with a prior anchored on cuff pressure, because UK Biobank has no invasive pressure. This is the largest weakness of the second stage, and posterior widths will tell us honestly whether it is fatal.
Baselines
The naive reference point is the common practice. It uses penalty displacement elements with the decoupled Holzapfel-Ogden energy, run with the same emulator design size, the same geometry and the same loading as every other cell. Beating it in stage one means that at least one shear-type parameter's total Sobol index exceeds a pre-set threshold that the naive cell does not reach. We set that threshold at the level the sensitivity study used to call the four main parameters influential. The naive cell must also reproduce the original finding that $a, b, a_f, b_f$ dominate; if it does not, our replication itself is suspect.
The strongest numerical competitor is the stabilized P1-P1 formulation of Karabelas et al. with the decoupled energy. It already removes locking and the split artifact, so it represents the reviewer's position that better elements alone suffice. The file reports its cost but no accuracy number usable here: on coarse artery meshes, projection-stabilized elements were about 3 times slower and MINI about 7 times slower than penalty formulations. We will run every cell to a fixed mesh-convergence tolerance in the end-systolic fiber stress rather than at fixed mesh size, so slower elements are not penalized by being run coarse. The coupled-versus-decoupled comparison within this formulation is the ablation that isolates the fabric ingredient.
For stage two, the strongest competitor is the ExWAS-style analysis itself. It uses raw regional peak circumferential strain as the phenotype, and the full set of MRI strain and shear features as a richer variant, in the same matched sample with the same covariates. Effect sizes for the TTN–strain association are summarized in the file only as counts: 224 gene-based burden associations across all MRI modalities, and 93.75% directional replication. There is no reported result on Lacuna for the TTN–circumferential strain effect size. We therefore re-estimate it on our sample. Beating it means a higher out-of-sample separation of carriers from matched non-carriers, measured as area under the ROC curve in nested cross-validation. Our fitted phenotype must also show a smaller dependence on cuff blood pressure than raw strain does. A third reference is sequence-only scoring by CPT-1 from Cross-protein transfer learning substantially improves disease variant prediction. CPT-1 reports 68% specificity at 95% ClinVar sensitivity, against 27% for ESM-1v and 55% for EVE. We do not compete with it as a classifier. Instead we test whether, among TTN missense carriers, the fitted parameter shift tracks the CPT-1 score. Agreement would support mechanical relevance, and disagreement would point to specific variants for follow-up.
Why not just…?
Why not just feed raw MRI strain into the association test, as ExWAS already does? This is the real threat. Raw strain has already detected TTN effects in Mapping rare protein-coding variants on multi-organ imaging traits, and it costs nothing extra. Our only argument for mechanics is that it removes loading and geometry confounds. We test exactly that by comparing ROC AUC and blood-pressure dependence head to head. If the fitted parameter does not beat a richer strain-feature set with blood pressure as a covariate, we will conclude that mechanics adds nothing to TTN association, and stage one stands alone as the contribution.
Why not just use locking-free elements and keep the standard energy? It may suffice. That is precisely the stabilized, $\kappa=0$ cell of the factorial design. If shear indices become large there, coupling is unnecessary and we say so. The textile evidence of Modeling multiaxial stress states in forming simulation of woven fabrics only suggests coupling matters under constrained, tension-loaded deformation. It does not prove it for myocardium.
Why not just refine the mesh or use higher-order displacement elements? Refinement reduces locking slowly, and the spurious fiber stress from the volumetric split documented by Karabelas et al. does not vanish with refinement. Even so, we include a refined penalty cell at matched wall-clock cost to the stabilized cell. If it closes the gap, the cheaper fix wins.
Why not just rely on sequence predictors for titin variants? Predictors stop at a probability, and damaging-looking variants are common in healthy people, as A calcium-sensing receptor allelic series and underdiagnosis of genetically driven hypocalcemia shows for another gene. Our proposal does not replace predictors. It asks whether a tissue-level measurement adds information about penetrance, and the CPT-1 concordance check reports that directly.
Decisive Checks
Three outcomes decide the project. The first is whether total Sobol indices of $a_{fs}, b_{fs}$ or $\kappa$ for shear-type strain observables cross the influence threshold in any locking-free cell while staying below it in the penalty cell. This tells us whether the published non-identifiability is partly a discretization artifact. The second is whether posterior intervals for the shear parameter on UK Biobank subjects are meaningfully narrower than the prior despite uncertain pressure. If they are not, stage two is unanswerable with this data and we stop. The third is whether, in nested cross-validation on matched TTN carriers and non-carriers, the fitted parameter beats the best raw-strain feature set on ROC AUC while depending less on cuff pressure. Positive answers to all three would justify mechanics-derived phenotypes for structural-gene variants. A negative first answer would still settle a modeling question that the sensitivity literature currently leaves open.
References
- Sensitivity analysis and inverse uncertainty quantification for the left ventricular passive mechanics
- Material Modelling of Fabric Deformation in Forming Simulation of Fiber-Metal Laminates – A Review on Modelling Fabric Coupling Mechanisms
- Modeling multiaxial stress states in forming simulation of woven fabrics
- Investigation of the Membrane Behavior of UD-NCF in Macroscopic Forming Simulations
- An accurate, robust, and efficient finite element framework with applications to anisotropic, nearly and fully incompressible elasticity
- Membrane behavior of uni- and bidirectional non-crimp fabrics in off-axis-tension tests
- Origin of tensile strength of a woven sample cut in bias directions
- Mapping rare protein-coding variants on multi-organ imaging traits
- Automated Quality Control in Image Segmentation: Application to the UK Biobank Cardiac MR Imaging Study
- Strain-Based Regional Nonlinear Cardiac Material Properties Estimation from Medical Images
- Cross-protein transfer learning substantially improves disease variant prediction
- A calcium-sensing receptor allelic series and underdiagnosis of genetically driven hypocalcemia