Pith. sign in

REVIEW 4 major objections 6 minor 71 references

A fully differentiable heart-wave simulator recovers hidden states and parameters from voltage-only observations, forecasting chaotic spiral dynamics 20-30 rotations ahead after learning only 4-5 rotations.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 23:11 UTC pith:ZFRMSEJF

load-bearing objection Serious methods paper with strong in-silico evidence for chaotic reentrant dynamics, but the abstract overclaims and the voltage-only identifiability question is never addressed. the 4 major comments →

arxiv 2607.15492 v1 pith:ZFRMSEJF submitted 2026-07-16 physics.med-ph nlin.CDphysics.bio-ph

Differentiable Cardiac Electrophysiology Simulations for Dynamical State and Parameter Estimation

Adarsh Pashikanti , Shrey Chowdhary , Alex Ho , Weizhen Li , Emilia Entcheva , Jan Christoph This is my paper
classification physics.med-ph nlin.CDphysics.bio-ph MSC 92C3035K5765M0665M75
keywords differentiable simulationscardiac electrophysiologystate estimationparameter estimationspiral wavesscroll wavesreaction-diffusiondigital twin
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that making cardiac electrophysiology simulations fully differentiable lets gradients from a simple voltage-mismatch loss flow back through the PDE solver, jointly recovering unknown model parameters and the complete initial dynamical state, including the hidden refractory variable. With a rolling schedule of short, overlapping training windows, the learned simulation co-evolves with the ground truth for tens of spiral rotations even though the dynamics are chaotic. The same machinery handles sparse electrode-grid data, surface-only observations of 3D scroll waves, and experimental optical-mapping recordings when augmented with a perceptual loss and a diffusion-model-based initialization. If correct, this offers a concrete route to personalizing heart models from the partial, noisy measurements available in clinical and experimental settings.

Core claim

The central claim is that gradient descent through a differentiable reaction-diffusion solver can solve the joint state-and-parameter estimation problem for cardiac tissue: observing only the voltage variable v on part of the domain over time, the method recovers the six parameters of a two-variable phenomenological model and the full initial (v0,r0) fields, with parameter errors under 0.05% for 2D spirals, 0.003-0.3% for 3D scroll waves, and 1-11% for focal or sparse-data cases. The strongest demonstration is long-horizon forecasting: after learning about 4-5 rotations of spiral-wave dynamics, the fitted simulation and the ground-truth system remain congruent for 20-30 rotations into the fu

What carries the argument

The load-bearing mechanism is an end-to-end differentiable PDE solver: the entire finite-difference or smoothed-particle-hydrodynamics integration is written in a way that permits exact reverse-mode gradients of the observation loss with respect to both the model parameters and the initial state. Around this, the decisive component is the rolling multi-horizon schedule - many short (10-40 time-step) overlapping horizons shifted by one time step - which prevents loss saturation and progressively refines the state and parameters; single long horizons plateau. A smoothness penalty on the hidden refractory field enables sparse-electrode recovery, and for experimental data a perceptual loss compu

Load-bearing premise

The entire recovery rests on the assumption that observing only the voltage variable over time uniquely determines the hidden refractory state, the initial condition, and the six model parameters - an identifiability condition the paper uses but never proves, and whose failure it exhibits locally in the biventricular focal case, where the recovered intramural pattern is called a degenerate solution.

What would settle it

A concrete check: construct two different (parameter, initial-state) configurations that produce identical voltage observations at all observed points and times (or differing below noise level) while having materially different hidden refractory fields; if such colliding configurations exist, the recovery is provably non-unique. A simpler experiment is to run the biventricular focal fitting with several random initializations and check whether the septal focus location and intramural pattern remain stable or vary across runs, as the paper already reports one distorted septal focus.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, clinically relevant hidden quantities become estimable: intramural early-activation sites and transmural reentrant wave patterns can be reconstructed from epicardial or dual-surface optical mapping.
  • Sparse multi-electrode-array recordings would suffice to reconstruct full-resolution reentrant dynamics, going beyond activation-time maps to the complete voltage and refractory fields.
  • Since chaotic spiral dynamics are forecast 20-30 rotations after only 4-5 rotations of observation, the approach could support prediction of arrhythmia evolution rather than mere diagnosis.
  • The multi-horizon schedule is a general recipe: it enables convergence even when initial parameter guesses are off by 50-100%, whereas single-horizon fitting saturates.
  • Model-to-model fitting suggests the framework can test how well candidate biophysical models reproduce observed dynamics, and the experimental fits point to a practical pipeline for real imaging data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The success on chaotic spirals suggests the method leverages the system's sensitive dependence: the chaotic attractor's dense sampling provides rich gradient information, implying the method is strongest precisely for the complex arrhythmic regimes where clinical prediction is hardest, and weakest for simple focal waves - consistent with the paper's own error trends.
  • If identifiability holds, the recovered refractory field constitutes a full tomographic reconstruction of transmural electrical state from surface data; a direct test would be to withhold some intramural electrode recordings from the fit and predict them.
  • The observed degeneration of intramural foci in the biventricular case indicates that observation time, not just spatial resolution, is a fundamental resource; translating to patients may require longer recordings or multiple episodes to disambiguate the hidden state.
  • A natural extension is to probe models with more state variables (e.g., ionic models); the method's identifiability with respect to hidden variables would likely degrade as model dimension grows, suggesting a trade-off between biophysical detail and recoverability.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces a fully differentiable cardiac electrophysiology simulation framework, implemented in JAX for finite-difference (2D/3D regular grids) and smoothed-particle-hydrodynamics (SPH) for heart-shaped geometries, and uses gradient-based optimization to estimate model parameters and the dynamical state from spatio-temporal voltage observations. The central method is a rolling multi-horizon learning schedule that sequentially fits overlapping time windows, which the authors show is critical for avoiding saturation and achieving parameter convergence. The claims are supported by in-silico experiments: recovery of all six Aliev-Panfilov parameters to <0.05–0.3% error for 2D spiral and 3D scroll waves, with out-of-sample co-evolution for tens of rotations; recovery of 2D multi-spiral dynamics from sparse electrode-grid observations; localization of intramural focal sources in a biventricular geometry from epicardial observations; and a cross-model fit of the AP model to Mitchell-Schaeffer data. For experimental monolayer imaging data, the authors augment the pixel-wise loss with a V-JEPA perceptual loss and use a diffusion model for initialization/constraint, showing fits to calcium and voltage spirals that remain visually congruent for several rotations. The paper is a proof-of-principle rather than a clinical tool, and it emphasizes that the method is most effective for complex chaotic reentrant dynamics.

Significance. If the quantitative recovery results hold, this is a substantial advance for state and parameter estimation in cardiac electrophysiology. The strongest evidence is the out-of-sample co-evolution test: after learning from about 4 spiral rotations, the fitted simulation and ground truth remain congruent for 20–30 rotation periods, and similar long-horizon agreement is demonstrated for 3D scroll waves from surface-only observations. The paper also introduces useful practical ingredients: the rolling multi-horizon schedule, a V-JEPA perceptual loss for noisy experimental data, and a DDPM-based initialization that projects initial states onto the model manifold. The SPH implementation on a realistic ventricular geometry extends differentiable-physics fitting beyond regular grids. The main weakness is that the identifiability premise—recovering hidden state and parameters from voltage-only, possibly partial observations—is asserted but not analyzed, and the paper itself documents a degeneracy in the biventricular focal case. This limits the transferability of the quantitative success to the experimental and clinical regimes.

major comments (4)
  1. [Section II C and III C] The paper's central premise is that observing only the voltage variable v on a subset of space-time determines the hidden state (v0, r0) and all parameters θ. This identifiability assumption is introduced in Section II C and never analyzed. The paper itself provides a counterexample in Section III C: after single-horizon fitting to epicardial observations, the epicardial loss is negligible while the intramural pattern 'might be a degenerate solution' and the septal focus is distorted. The abstract's broad claim—'recover the full dynamics, even with sparse, noisy, or partial observations'—is therefore not supported. The authors should either characterize the conditions under which the inverse problem is identifiable (e.g., via observability analysis or a parameter-sensitivity study) or restrict the claim to chaotic reentrant dynamics, which are empirically more informative.
  2. [Section II C 3 and III B] In the rolling multi-horizon schedule, each new horizon re-initializes the state using the second learned state of the previous horizon and the observed voltage at the new horizon boundary (Eq. 11 and surrounding text). This periodically re-anchors the simulation to ground-truth observations during training, so the parameter gradients are accumulated across teacher-forced windows rather than a single free-running trajectory. The out-of-sample forecast is a valid control, but the paper does not report an ablation in which the learned parameters/state are evaluated without such re-anchoring during training. This makes it difficult to attribute the parameter recovery accuracy to the differentiable solver alone versus the re-anchoring schedule. A control experiment (e.g., fitting with only the first horizon and then forecasting) would clarify the mechanism.
  3. [Section III G, Fig. 17] The experimental fits are evaluated qualitatively: the text states that the simulated spiral is 'slightly slower' or 'slightly faster' and that the two co-evolve for 'about 4–5 rotations' before divergence, but no quantitative error metric, action-potential-duration comparison, conduction-velocity comparison, or phase-singularity-trajectory error is provided. Given the paper's claim that the method transfers to real experimental footage, a quantitative assessment of the fit quality (even a simple pixel-wise or phase-based error over the co-evolution window) is needed. Without it, the experimental section remains a visual demonstration rather than a measured validation.
  4. [Section III A and IV] The paper reports that parameter errors for focal waves are 1–5% for {ε0, μ1, μ2} (Table IV) and states in Section IV that learning is less effective for focal or stationary waves. This is a limitation of the method for an important class of cardiac rhythms (pacing, premature beats). The authors should state this limitation prominently in the abstract or introduction, since the current abstract implies universal applicability. The technical explanation offered—that complex dynamics are more informative—is plausible but not tested; a brief analysis or a more cautious wording would improve the paper.
minor comments (6)
  1. [Eq. (6)] The regularization loss uses the symbol u in max(-u,0) but the state variables are v and r; presumably u is a typo for v. Please clarify.
  2. [Fig. 5B] The caption text is corrupted: 'With 27 horizons, the dynamics are identical over 20 rotations. or 67 time steps or 20,000 epochs, the dynamics is learned overCTop row:' contains a truncated sentence. Please rewrite the caption.
  3. [Section II C 1, Eq. (9)] The perceptual loss Lp is written as a sum over t of a quotient. The notation for the ℓ2-normalized features is ambiguous; please define N or use explicit vector norms for each frame, e.g., ||fθ(v't)||2.
  4. [Section II C 3] The phrase 'with strides=1' appears to be a typo; the intended term is 'stride'. Also, the construction (v',r')_{i-1,2} → (v_bar,r')_{i,1} is not fully explained in the main text; consider a diagram or a more explicit definition.
  5. [Section II C 4] The description of the DDPM initialization states that the model can translate MS voltage patterns into AP spirals with both components, but the SDEdit/RePaint procedure is only outlined. Since this is a key component for the experimental fits, a more detailed description in the supplement would be helpful.
  6. [General] The source code is stated to be available upon publication, but no repository link is provided in the manuscript. Please add a data/code availability statement with a URL.

Circularity Check

0 steps flagged

No significant circularity; the central inverse problem and long-term forecasts are not defined by their fit inputs.

full rationale

Most claims are tested against quantities outside the fitting objective. In the in-silico experiments, the observation loss (Eq. 5) is computed only over the voltage variable v within each horizon; the hidden refractory field r, the initial condition (v0,r0), and the parameters theta are all free variables. The long-term 'forecasts' (Figs. 5, 6, 14) are obtained by evolving the learned state beyond the last training horizon and comparing to ground truth; this future interval is never used in backpropagation, so the co-evolution result is not a fitted value renamed as a prediction. The experimental section is anchored to real optical/calcium imaging video, an external benchmark; the V-JEPA perceptual loss and DDPM are fixed regularizers/initializers and are not fitted to the experimental evaluation frames. The paper even exhibits the opposite of circular success in Section III C: the epicardial loss becomes negligible while the intramural solution 'might be a degenerate solution,' showing the method does not force the hidden state to equal the input by construction. The self-citations (Baranwal et al. [50] for the DDPM, Chowdhary et al. [2] for panoramic imaging) are methodological/data references and are not load-bearing for the central state-and-parameter estimation claim. The main caveat is that identifiability from voltage-only, partially observed data is assumed rather than proven; that is a completeness/correctness risk, not a circular step.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The central recovery claims rest on the assumption that observing v on a partial grid is informative enough to determine the hidden state and parameters; no identifiability proof is offered, and the paper itself surfaces a degenerate solution in the biventricular focal case. The AP model is taken as the generative model for the experimental data, which is plausible but not quantitatively demonstrated. The discretizations are standard numerical assumptions; the V-JEPA encoder and DDPM are ad hoc components trained on the same model class as the fitted target.

free parameters (6)
  • smoothness loss weight alpha = 0.01
    Chosen by hand in Section II C 1 for sparse-data experiments; with alpha=0 learning fails, with larger values dynamics are suppressed.
  • perceptual loss weight w_p = 2.0
    Set manually in Section III G for the experimental fit; no sensitivity analysis is reported.
  • learning rates = 1e-3 (FD), 1e-4 (SPH); constant or exponential decay
    Chosen empirically; the paper notes that constant vs. decay schedules matter (Fig. S3D).
  • multi-horizon schedule hyperparameters = Tables II-III: 3 stages of 10 horizons with durations 10/20/40 and epochs 600/800/1000; 4 warm-up plus 51 rolling horizo
    Load-bearing: 'we empirically found that multi-horizon learning outperforms learning the dynamics all at once' (II C 3); exact durations, overlaps, and epoch counts are hand-tuned.
  • AP parameters fitted to experimental data = not reported
    For the two monolayer fits (III G), the six AP parameters are fitted to data, but final values, uncertainties, and identifiability are not reported.
  • Gaussian noise sigma in r'_0 initialization = unspecified
    Equation (12) sets r'_0 = v_0(1+sigma); the noise amplitude is not reported, and the initialization is stated to strongly affect learning (Fig. S1, S3A).
axioms (5)
  • domain assumption Observing only the voltage variable v at sampled points/times is sufficient to identify v0, r0 and theta up to the reported accuracy.
    Used in all experiments (Section II C and II C 1); no identifiability or observability analysis is given; degeneracy is acknowledged in III C.
  • domain assumption The Aliev-Panfilov model (equations 1-2) is an adequate generative model for the observed experimental calcium/voltage spirals.
    Stated for the experimental fits (III G); the paper itself notes AP-MSmismatch leaves residual errors (III F), so this assumption is non-trivial.
  • standard math The finite-difference (9-point Laplacian, Tsit5, tolerances 1e-4) and SPH (Wendland C2, h=1.2Dx, Strang splitting, QSS integrator) discretizations faithfully represent the PDEs for the purpose of gradient-based optimization.
    Section II B; discretization convergence of gradients is assumed, not proven.
  • ad hoc to paper The V-JEPA encoder pre-trained on AP-model spiral videos provides a perceptual loss that guides the fit toward the true experimental dynamics.
    Section II C 1 and III G; the encoder is trained on synthetic AP spirals produced by the same group, so the loss measures 'AP-likeness' rather than an independent perceptual metric.
  • ad hoc to paper The DDPM trained on AP simulation data can translate observed voltage patterns into valid AP initial conditions with both state variables.
    Section II C 4; used for cross-model and experimental fits; no quantitative validation of the translated initial states is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 30574 in / 17803 out tokens · 185907 ms · 2026-08-01T23:11:03.983176+00:00 · methodology

0 comments
read the original abstract

The heart's contractions are triggered by action potential waves, which propagate through the cardiac muscle and exhibit diverse spatio-temporal dynamics during different heart rhythms. The dynamics are modeled with partial differential equations (PDEs) in cardiac electrophysiology simulations. However, fitting such models to measurement data to develop digital twins or patient-specific computer models is challenging. Here, we introduce differentiable cardiac electrophysiology simulations that can be fitted automatically to spatio-temporal measurement data of action potential waves in cardiac tissue. By comparing the simulated dynamics with the observation data, we define a loss function that is minimized via gradient-based optimization. Backpropagating the loss gradient through the differentiable PDE solver enables us to learn the parameters and recover the full dynamics, even with sparse, noisy, or partial observations. Implemented using both the finite-difference and smoothed particle hydrodynamics methods, our simulation framework can be applied to pixel-, voxel-, or point-based data, such as 2D or 3D slabs, or arbitrary shapes, such as the heart's ventricles. Using this methodology, we locate early activation sites inside a 3D bi-ventricular simulation geometry and fit a phenomenological model to imaging data of a voltage spiral wave in a cardiac monolayer cell culture. With experimental data, we employed a perceptual loss based on the Video Joint-Embedding Predictive Architecture, which enables fitting to noisy imaging data, and a generative diffusion model to estimate initial conditions and constrain solutions. Differentiable cardiac electrophysiology simulations could improve the diagnosis of rhythm abnormalities in patients and facilitate the development of personalized models or digital twins of the heart.

Figures

Figures reproduced from arXiv: 2607.15492 by Adarsh Pashikanti, Alex Ho, Emilia Entcheva, Jan Christoph, Shrey Chowdhary, Weizhen Li.

Figure 1
Figure 1. Figure 1: FIG. 1. Gradient-based fitting to spatio-temporal data using differen [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. Simulations of electrical action potential waves (red: de [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Bi-ventricular simulations of cardiac electrophysiology us [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. Learning the initial condition of a focal wave during the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. Reconstruction of spiral wave dynamics using multi-horizon learning schedule with a differentiable simulation. Effect of observation [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. Agreement between learned and original ground-truth (GT) spiral wave dynamics after gradient-descent-based learning with different [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7. Convergence of learned Aliev Panfilov model parameters [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8. Learning and predicting meandering spiral wave dynamics. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9. Learning of transmural bi-focal action potential wave pattern from epicardial observations using differentiable smoothed particle [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: FIG. 10. Localization of endocardial early activation sites from epicardial action potential wave dynamics. [PITH_FULL_IMAGE:figures/full_fig_p014_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: FIG. 11. Recovery of 3D electrical scroll wave from epicardial obser [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 14
Figure 14. Figure 14: FIG. 14. Learning of scroll wave dynamics by observing 3-4 rota [PITH_FULL_IMAGE:figures/full_fig_p016_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: FIG. 15. State and parameter recovery with sparse observations. [PITH_FULL_IMAGE:figures/full_fig_p017_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: FIG. 16. Cross-model learning: fitting the Aliev-Panfilov (AP) [PITH_FULL_IMAGE:figures/full_fig_p018_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: FIG. 17. Automated fitting to a spiral wave in two different mono [PITH_FULL_IMAGE:figures/full_fig_p019_17.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

71 extracted references · 4 canonical work pages

  1. [1]

    Gepstein, G

    L. Gepstein, G. Hayam, and S. A. Ben-Haim, A novel method for nonfluoroscopic catheter-based electroanatomical mapping of the heart, Circulation95, 1611 (1997)

    1997

  2. [2]

    Chowdhary, J

    S. Chowdhary, J. Lebert, S. Dickman, M. Manetta, C. Gor- don, and J. Christoph, Panoramic voltage-sensitive optical map- ping of contracting hearts using cooperative multiview mo- tion tracking with 12 cameras, The Journal of Physiologyn/a, https://doi.org/10.1113/JP290533 (2026)

    work page doi:10.1113/jp290533 2026

  3. [3]

    J. M. Davidenko, A. V . Pertsov, R. Salomonsz, W. Baxter, and J. Jalife, Stationary and drifting spiral waves of excitation in isolated cardiac muscle, Nature355, 349 (1992)

    1992

  4. [4]

    A. M. Pertsov, R. Davidenko, J. M. Salomonsz, W. T. Baxter, and J. Jalife, Spiral waves of excitation underlie reentrant ac- tivity in isolated cardiac muscle, Circulation Research72, 631 (1993)

    1993

  5. [5]

    Alonso, M

    S. Alonso, M. Bär, and B. Echebarria, Nonlinear physics of electrical wave propagation in the heart: a review, Reports on Progress in Physics79, 096601 (2016)

    2016

  6. [6]

    Rappel, The physics of heart rhythm disorders, Physics Reports978, 1 (2022)

    W.-J. Rappel, The physics of heart rhythm disorders, Physics Reports978, 1 (2022). 22

    2022

  7. [7]

    Christoph, M

    J. Christoph, M. Chebbok, C. Richter, J. Schröder-Schetelig, P. Bittihn, S. Stein, I. Uzelac, F. H. Fenton, G. Hasenfuss, R. J. Gilmour, and S. Luther, Electromechanical vortex fila- ments during cardiac fibrillation, Nature555, 667 (2018)

    2018

  8. [8]

    S. Berg, S. Luther, and U. Parlitz, Synchronization based system identification of an extended excitable system, Chaos: An Interdisciplinary Journal of Nonlinear Science21, 10.1063/1.3613921 (2011)

    work page doi:10.1063/1.3613921 2011

  9. [9]

    M. J. Hoffman, N. S. LaVigne, S. T. Scorse, F. H. Fenton, and E. M. Cherry, Reconstructing three-dimensional reentrant car- diac electrical wave dynamics using data assimilation, Chaos: An Interdisciplinary Journal of Nonlinear Science26, 013107 (2016)

    2016

  10. [10]

    N. S. LaVigne, N. Holt, M. J. Hoffman, and E. M. Cherry, Ef- fects of model error on cardiac electrical wave state reconstruc- tion using data assimilation, Chaos: An Interdisciplinary Jour- nal of Nonlinear Science27, 093911 (2017)

    2017

  11. [11]

    M. J. Hoffman and E. M. Cherry, Sensitivity of a data- assimilation system for reconstructing three-dimensional car- diac electrical dynamics, Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sci- ences378, 20190388 (2020)

    2020

  12. [12]

    C. D. Marcotte, F. H. Fenton, M. J. Hoffman, and E. M. Cherry, Robust data assimilation with noise: Applications to cardiac dy- namics, Chaos: An Interdisciplinary Journal of Nonlinear Sci- ence31, 013118 (2021)

    2021

  13. [13]

    C. D. Marcotte, M. J. Hoffman, F. H. Fenton, and E. M. Cherry, Reconstructing cardiac electrical excitations from optical map- ping recordings, Chaos: An Interdisciplinary Journal of Non- linear Science33, 093141 (2023)

    2023

  14. [14]

    M. J. Mendez, E. M. Cherry, G. S. Hoeker, S. Poelzing, and S. H. Weinberg, Reconstructing ventricular cardiomyocyte dy- namics and parameter estimation using data assimilation, Bio- physical Journal123, 4050 (2024)

    2024

  15. [15]

    Seemann, S

    G. Seemann, S. Lurz, D. U. J. Keller, D. L. Weiss, E. P. Scholz, and O. Dössel, Adaption of mathematical ion channel models to measured data using the particle swarm optimization, in4th European Conference of the International Federation for Med- ical and Biological Engineering, edited by J. Vander Sloten, P. Verdonck, M. Nyssen, and J. Haueisen (Springer...

    2009

  16. [16]

    F. Chen, A. Chu, X. Yang, Y . Lei, and J. Chu, Identification of the parameters of the beeler–reuter ionic equation with a partially perturbed particle swarm optimization, IEEE Trans- actions on Biomedical Engineering59, 3412 (2012)

    2012

  17. [17]

    Loewe, M

    A. Loewe, M. Wilhelms, J. Schmid, M. J. Krause, F. Fischer, D. Thomas, E. P. Scholz, O. Dössel, and G. Seemann, Pa- rameter estimation of ion current formulations requires hybrid optimization approach to be both accurate and reliable, Fron- tiers in Bioengineering and BiotechnologyV olume 3 - 2015, 10.3389/fbioe.2015.00209 (2016)

    arXiv 2015

  18. [18]

    Rheaume, H

    E. Rheaume, H. Velasco-Perez, D. Cairns, M. Comstock, E. Rheaume, A. Kaboudian, I. Uzelac, E. Cherry, and F. H. Fenton, A modified fitzhugh-nagumo model that reproduces the action potential and dynamics of the ten tusscher et al. cardiac model in tissue, Computing in Cardiology 202350(2023)

    2023

  19. [19]

    D. I. Cairns, M. R. Comstock, F. H. Fenton, and E. M. Cherry, Cardiofit: a webgl-based tool for fast and efficient parametriza- tion of cardiac action potential models to fit user-provided data, Royal Society Open Science12, 250048 (2025)

    2025

  20. [20]

    Dokos and N

    S. Dokos and N. H. Lovell, Parameter estimation in cardiac ionic models, Progress in Biophysics and Molecular Biology 85, 407 (2004), modelling Cellular and Tissue Function

    2004

  21. [21]

    D. M. Lombardo, F. H. Fenton, S. M. Narayan, and W.-J. Rap- pel, Comparison of detailed and simplified models of human atrial myocytes to recapitulate patient specific properties, PLOS Computational Biology12, 1 (2016)

    2016

  22. [22]

    Z. Syed, E. Vigmond, S. Nattel, and L. J. Leon, Atrial cell ac- tion potential parameter fitting using genetic algorithms, Medi- cal and Biological Engineering and Computing43, 561 (2005)

    2005

  23. [23]

    C. T. Bot, A. R. Kherlopian, F. A. Ortega, D. J. Christini, and T. Krogh-Madsen, Rapid genetic algorithm optimization of a mouse computational model: Benefits for anthropomorphiza- tion of neonatal mouse cardiomyocytes, Frontiers in Physiology V olume 3 - 2012, 10.3389/fphys.2012.00421 (2012)

    Pith/arXiv arXiv 2012

  24. [24]

    D. I. Cairns, F. H. Fenton, and E. M. Cherry, Efficient parame- terization of cardiac action potential models using a genetic al- gorithm, Chaos: An Interdisciplinary Journal of Nonlinear Sci- ence27, 093922 (2017)

    2017

  25. [25]

    C. H. Martin, A. Oved, R. A. Chowdhury, E. Ullmann, N. S. Peters, A. A. Bharath, and M. Varela, EP-PINNs: Car- diac electrophysiology characterisation using physics-informed neural networks, Frontiers in Cardiovascular Medicine8, 10.3389/fcvm.2021.768419 (2022)

    arXiv 2021

  26. [26]

    Pezzuto, P

    S. Pezzuto, P. Perdikaris, and F. S. Costabal, Learning car- diac activation maps from 12-lead ecg with multi-fidelity bayesian optimization on manifolds, IFAC-PapersOnLine55, 175 (2022), 10th Vienna International Conference on Mathe- matical Modelling MATHMOD 2022

    2022

  27. [27]

    A. Neic, F. O. Campos, A. J. Prassl, S. A. Niederer, M. J. Bishop, E. J. Vigmond, and G. Plank, Efficient computation of electrograms and ecgs in human whole heart simulations us- ing a reaction-eikonal model, Journal of Computational Physics 346, 191 (2017)

    2017

  28. [28]

    Pezzuto, F

    S. Pezzuto, F. W. Prinzen, M. Potse, F. Maffessanti, F. Regoli, M. L. Caputo, G. Conte, R. Krause, and A. Auricchio, Reconstruction of three-dimensional biven- tricular activation based on the 12-lead electrocardio- gram via patient-specific modelling, EP Europace23, 640 (2020), https://academic.oup.com/europace/article- pdf/23/4/640/36916991/euaa330.pdf

    2020

  29. [29]

    Gillette, M

    K. Gillette, M. A. Gsell, A. J. Prassl, E. Karabelas, U. Reiter, G. Reiter, T. Grandits, C. Payer, D. Štern, M. Urschler, J. D. Bayer, C. M. Augustin, A. Neic, T. Pock, E. J. Vigmond, and G. Plank, A framework for the generation of digital twins of cardiac electrophysiology from clinical 12-leads ecgs, Medical Image Analysis71, 102080 (2021)

    2021

  30. [30]

    Camps, Z

    J. Camps, Z. J. Wang, R. Doste, L. A. Berg, M. Holmes, B. Lawson, J. Tomek, K. Burrage, A. Bueno-Orovio, and B. Rodriguez, Harnessing 12-lead ecg and mri data to person- alise repolarisation profiles in cardiac digital twin models for enhanced virtual drug testing, Medical Image Analysis100, 103361 (2025)

    2025

  31. [31]

    Doste, J

    R. Doste, J. Camps, Z. J. Wang, L. A. Berg, M. Holmes, H. Smith, M. Beetz, L. Li, A. Banerjee, V . Grau, and B. Ro- driguez, An automated computational pipeline for generating large-scale cohorts of patient-specific ventricular models in electromechanical in silico trials, Computer Methods and Pro- grams in Biomedicine279, 109290 (2026)

    2026

  32. [32]

    Y . Hu, L. Anderson, T.-M. Li, Q. Sun, N. Carr, J. Ragan- Kelley, and F. Durand, Difftaichi: Differentiable programming for physical simulation (2020), arXiv:1910.00935 [cs.LG]

    Pith/arXiv arXiv 2020

  33. [33]

    Thuerey, B

    N. Thuerey, B. Holzschuh, P. Holl, G. Kohl, M. Lino, Q. Liu, P. Schnell, and F. Trost, Physics-based deep learning (2025), arXiv:2109.05237 [cs.LG]

    Pith/arXiv arXiv 2025

  34. [34]

    Bradbury, R

    J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. VanderPlas, 23 S. Wanderman-Milne, and Q. Zhang, Jax: Composable trans- formations of python+numpy programs (2018)

    2018

  35. [35]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Köpf, E. Yang, Z. DeVito, M. Raison, A. Te- jani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chin- tala, Pytorch: an imperative style, high-performance deep learn- ing library, inProceedings of the 33rd International Co...

    2019

  36. [36]

    Winchenbach and N

    R. Winchenbach and N. Thuerey, diffsph: Differentiable smoothed particle hydrodynamics for hybrid machine learning solutions in fluid mechanics, Journal of Computational Physics 555, 114769 (2026)

    2026

  37. [37]

    S. S. Schoenholz and E. D. Cubuk, Jax, m.d. a framework for differentiable physics, inProceedings of the 34th International Conference on Neural Information Processing Systems, NIPS ’20 (Curran Associates Inc., Red Hook, NY , USA, 2020)

    2020

  38. [38]

    B. J. Thomas, C. Goodbrake, K. Meyer, and M. S. Sacks, High speed cardiac simulations using the jax framework, inFunc- tional Imaging and Modeling of the Heart, edited by R. Chabin- iok, Q. Zou, T. Hussain, H. H. Nguyen, V . G. Zaha, and M. Gus- seva (Springer Nature Switzerland, Cham, 2025) pp. 275–281

    2025

  39. [39]

    Lettermann, A

    L. Lettermann, A. Jurado, T. Betz, F. Wörgötter, and S. Her- zog, Tutorial: a beginner’s guide to building a representative model of dynamical systems using the adjoint method, Com- mun. Phys.7(2024)

    2024

  40. [40]

    Kashtanova, M

    V . Kashtanova, M. Pop, I. Ayed, P. Gallinari, and M. Serme- sant, Simultaneous data assimilation and cardiac electrophys- iology model correction using differentiable physics and deep learning, Interface Focus13, 20230043 (2023)

    2023

  41. [41]

    R. R. Aliev and A. V . Panfilov, A simple two-variable model of cardiac excitation, Chaos, Solitons & Fractals7, 293 (1996)

    1996

  42. [42]

    C. C. Mitchell and D. G. Schaeffer, A two-current model for the dynamics of cardiac membrane, Bulletin of Mathematical Biology65, 767 (2003)

    2003

  43. [43]

    Zhang, M

    C. Zhang, M. Rezavand, Y . Zhu, Y . Yu, D. Wu, W. Zhang, J. Wang, and X. Hu, SPHinXsys: An open-source multi-physics and multi-resolution library based on smoothed particle hy- drodynamics, Computer Physics Communications267, 108066 (2021)

    2021

  44. [44]

    Zhang, J

    C. Zhang, J. Wang, M. Rezavand, D. Wu, and X. Hu, An inte- grative smoothed particle hydrodynamics method for modeling cardiac function, Computer Methods in Applied Mechanics and Engineering381, 113847 (2021)

    2021

  45. [45]

    Kidger,On Neural Differential Equations, Ph.D

    P. Kidger,On Neural Differential Equations, Ph.D. thesis, Uni- versity of Oxford (2021)

    2021

  46. [46]

    Kidger and C

    P. Kidger and C. Garcia, Equinox: neural networks in jax via callable pytrees and filtered transformations (2021), arXiv:2111.00254 [cs.LG]

    Pith/arXiv arXiv 2021

  47. [47]

    D. P. Kingma and J. Ba, Adam: A method for stochastic opti- mization (2017), arXiv:1412.6980 [cs.LG]

    Pith/arXiv arXiv 2017

  48. [48]

    Doste, D

    R. Doste, D. Soto-Iglesias, G. Bernardino, A. Al- caine, R. Sebastian, S. Giffard-Roisin, M. Sermesant, A. Berruezo, D. Sanchez-Quintana, and O. Camara, A rule-based method to model myocardial fiber orientation in cardiac biventricular geometries with outflow tracts, International Journal for Numerical Methods in Biomed- ical Engineering35, e3185 (2019), ...

    doi:10.1002/cnm.3185 2019

  49. [49]

    L. Eing, C. Luna-Jiménez, S. Mertes, and E. André, Video joint- embedding predictive architectures for facial expression recog- nition (2026), arXiv:2601.09524 [cs.CV]

    arXiv 2026

  50. [50]

    Baranwal, J

    T. Baranwal, J. Lebert, and J. Christoph, Dreaming of electrical waves: Generative modeling of cardiac excitation waves using diffusion models, APL Machine Learning2, 036113 (2024)

    2024

  51. [51]

    C. Meng, Y . He, Y . Song, S. Jiaming, J. Wu, J.-Y . Zhu, and S. Ermon, Sdedit: Guided image synthesis and editing with stochastic differential equations, inInternational Conference on Learning Representations (ICLR)(2022)

    2022

  52. [52]

    Lugmayr, M

    A. Lugmayr, M. Danelljan, A. Romero, F. Yu, R. Timofte, and L. Van Gool, Repaint: Inpainting using denoising diffusion probabilistic models, inProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) (2022) pp. 11451–11461

    2022

  53. [53]

    Monteiro da Rocha, G

    A. Monteiro da Rocha, G. Guerrero-Serna, A. Helms, C. Luzod, S. Mironov, M. Russell, J. Jalife, S. M. Day, G. D. Smith, and T. J. Herron, Deficient cmybp-c protein expression during car- diomyocyte differentiation underlies human hypertrophic car- diomyopathy cellular phenotypes in disease specific human es cell derived cardiomyocytes, Journal of Molecula...

    2016

  54. [54]

    Y . W. Heinson, J. L. Han, and E. Entcheva, Portable low-cost macroscopic mapping system for all-optical cardiac electro- physiology, Journal of Biomedical Optics28, 016001 (2023)

    2023

  55. [55]

    W. Liu, J. L. Han, J. Tomek, G. Bub, and E. Entcheva, Simulta- neous widefield voltage and dye-free optical mapping quantifies electromechanical waves in human induced pluripotent stem cell-derived cardiomyocytes, ACS Photonics10, 1070 (2023), https://doi.org/10.1021/acsphotonics.2c01644

    work page doi:10.1021/acsphotonics.2c01644 2023

  56. [56]

    Huang, A

    Y .-L. Huang, A. S. Walker, and E. W. Miller, A photostable silicon rhodamine platform for optical voltage sensing, Journal of the American Chemical Society137, 10767—10776 (2015)

    2015

  57. [57]

    Klimas, G

    A. Klimas, G. Ortiz, S. C. Boggess, E. W. Miller, and E. Entcheva, Multimodal on-axis platform for all-optical elec- trophysiology with near-infrared probes in human stem-cell- derived cardiomyocytes, Progress in Biophysics and Molecu- lar Biology154, 62 (2020), novel optics-based approaches for cardiac electrophysiology

    2020

  58. [58]

    B. G. Mitrea, M. Wellner, and A. M. Pertsov, Monitoring in- tramyocardial reentry using alternating transillumination, in 2009 Annual International Conference of the IEEE Engineer- ing in Medicine and Biology Society(2009) pp. 4194–4197

    2009

  59. [59]

    Bueno-Orovio, E

    A. Bueno-Orovio, E. M. Cherry, and F. H. Fenton, Minimal model for human ventricular action potentials in tissue, Jour- nal of Theoretical Biology253, 544 (2008)

    2008

  60. [60]

    C.-E. Chiu, A. Roy, S. Cechnicka, A. Gupta, A. L. Pinto, C. Galazis, K. Christensen, D. Mandic, and M. Varela, Physics- informed neural networks can accurately model cardiac electro- physiology in 3d geometries and fibrillatory conditions, inSta- tistical Atlases and Computational Models of the Heart. Work- shop, CMRxRecon and MBAS Challenge Papers.(Sprin...

    2025

  61. [61]

    Lebert, M

    J. Lebert, M. Mittal, and J. Christoph, Reconstruction of three-dimensional scroll waves in excitable media from two- dimensional observations using deep neural networks, Phys. Rev. E107, 014221 (2023)

    2023

  62. [62]

    Stenger, S

    R. Stenger, S. Herzog, I. Kottlarz, B. Rüchardt, S. Luther, F. Wörgötter, and U. Parlitz, Reconstructing in-depth activity for chaotic 3d spatiotemporal excitable media models based on surface data, Chaos: An Interdisciplinary Journal of Nonlinear Science33, 013134 (2023), https://doi.org/10.1063/5.0126824

    work page doi:10.1063/5.0126824 2023

  63. [63]

    Eckstein, B

    J. Eckstein, B. Maesen, D. Linz, S. Zeemering, A. van Hun- nik, S. Verheule, M. Allessie, and U. Schotten, Time course and mechanisms of endo-epicardial electrical dissociation dur- ing atrial fibrillation in the goat, Cardiovascular Research89, 816 (2010). 24

    2010

  64. [64]

    de Groot, L

    N. de Groot, L. van der Does, A. Yaksh, E. Lanters, C. Teuwen, P. Knops, P. van de Woestijne, J. Bekkers, C. Kik, A. Bogers, and M. Allessie, Direct proof of endo-epicardial asynchrony of the atrial wall during atrial fibrillation in humans, Circulation: Arrhythmia and Electrophysiology9, e003648 (2016)

    2016

  65. [65]

    T. E. Walters, G. Lee, A. Lee, R. Sievers, J. M. Kalman, and E. P. Gerstenfeld, Site-specific epicardium-to-endocardium dis- sociation of electrical activation in a swine model of atrial fib- rillation, JACC: Clinical Electrophysiology6, 830 (2020)

    2020

  66. [66]

    Fenton and A

    F. Fenton and A. Karma, V ortex dynamics in three-dimensional continuous myocardium with fiber rotation: Filament instabil- ity and fibrillation, Chaos: An Interdisciplinary Journal of Non- linear Science8, 20 (1998)

    1998

  67. [67]

    Berenfeld and A

    O. Berenfeld and A. M. Pertsov, Dynamics of intramural scroll waves in three-dimensional continuous myocardium with ro- tational anisotropy, Journal of Theoretical Biology199, 383 (1999)

    1999

  68. [68]

    Z. Qu, J. Kil, F. Xie, A. Garfinkel, and J. N. Weiss, Scroll wave dynamics in a three-dimensional cardiac tissue model: Roles of restitution, thickness, and fiber rotation, Biophysical Journal 78, 2761 (2000)

    2000

  69. [69]

    Bernus, K

    O. Bernus, K. S. Mukund, and A. M. Pertsov, Detection of in- tramyocardial scroll waves using absorptive transillumination imaging, Journal of Biomedical Optics12, 014035 (2007). 25 SUPPLEMENTARY INFORMA TION Supplementary Videos Supplementary Videos will be provided at:https:// cardiacvision.ucsf.edu/videos/diffsim/. A. Supplementary Tables Param. D k aε...

    2007

  70. [70]

    The horizon du- ration increases from 10 to 20 to 40 time steps in each stage, and each horizon is shifted by 1 time step (∆t=1)

    The schedule comprises 3 subsequent stages with 10 horizons each and 600, 800, and 1,000 epochs per horizon. The horizon du- ration increases from 10 to 20 to 40 time steps in each stage, and each horizon is shifted by 1 time step (∆t=1). After 10, 20, and 30 horizons, the learning was performed over 20, 40, and 70 time steps, respectively. At the end of ...

  71. [71]

    The horizon duration increases during warm- up and then stays constant during the rolling horizons

    The schedule comprises 4 subsequent "warm-up" horizons (W1- W4), which refine the initial condition, followed by a rolling multi- horizon schedule comprising 51 horizons, which learn the dynamics, see also section III D. The horizon duration increases during warm- up and then stays constant during the rolling horizons. Each horizon is learned over 2,000 e...

    2000