Learning Cardiac Electrophysiology Digital Twins Through Agentic Discovery of Hybrid Structure

Linwei Wang; Sumeet Atul Vadhavka; Yubo Ye; Zhiqiang Tao; Ziqi Zhou

arxiv: 2606.18154 · v1 · pith:SEPCJZX3new · submitted 2026-06-16 · 💻 cs.AI

Learning Cardiac Electrophysiology Digital Twins Through Agentic Discovery of Hybrid Structure

Ziqi Zhou , Yubo Ye , Sumeet Atul Vadhavka , Linwei Wang , Zhiqiang Tao This is my paper

Pith reviewed 2026-06-27 00:46 UTC · model grok-4.3

classification 💻 cs.AI

keywords cardiac electrophysiologydigital twinshybrid modelsLLM agentmodel discoveryaction spacepersonalized modeling

0 comments

The pith

An LLM agent discovers stable hybrid models for patient-specific cardiac electrophysiology digital twins by treating domain knowledge as an action space.

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

The paper claims that building accurate digital twins for individual hearts requires discovering the right model structure rather than just tuning parameters in a fixed architecture. Existing expert-designed hybrids demand deep knowledge and fail to generalize across patients, while pure LLM approaches produce unstable simulations. LEADS gives the LLM a structured action space drawn from cardiac electrophysiology knowledge so the agent can iteratively select, combine, and refine physics-neural components; gradient descent then fits the parameters. Validation on synthetic data with known reaction models and on real cardiac recordings shows the discovered models outperform both hand-crafted hybrids and other LLM-based methods in stability and accuracy.

Core claim

LEADS formulates cardiac EP domain knowledge as a structured action space that an LLM agent traverses through an iterative reasoning-and-action loop. The agent selects, combines, and refines hybrid physics-neural structures while gradient descent optimizes parameters, producing models that remain physically grounded, interpretable, and numerically stable without requiring the end user to supply deep domain expertise.

What carries the argument

The structured action space of cardiac EP domain knowledge that lets the LLM agent perform iterative selection, combination, and refinement of hybrid models.

If this is right

Patient-specific digital twins become feasible without per-patient expert redesign of the model architecture.
The same agentic loop can be applied to any new patient recording to yield a tailored hybrid structure.
Discovered models remain interpretable because each component is drawn from the predefined action vocabulary.
Numerical stability during long simulations improves because the action space explicitly favors stable reaction terms.

Where Pith is reading between the lines

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

The approach could extend to other excitable-media systems such as neural tissue or chemical reaction networks if analogous action spaces are defined.
Once a library of discovered structures accumulates, meta-learning over the action histories might accelerate discovery for future patients.
The method assumes the action space already encodes the necessary priors; expanding or pruning that space would directly test how much domain knowledge is truly required.

Load-bearing premise

Encoding cardiac electrophysiology knowledge into a fixed set of allowable actions is enough for the LLM to generate models that remain physically valid and numerically stable across patients.

What would settle it

Run the discovered LEADS models on the held-out real cardiac EP recordings and check whether they produce unstable voltage traces or larger errors than the best human-designed hybrid baseline.

Figures

Figures reproduced from arXiv: 2606.18154 by Linwei Wang, Sumeet Atul Vadhavka, Yubo Ye, Zhiqiang Tao, Ziqi Zhou.

**Figure 1.** Figure 1: Overview of LEADS. Left: the LLM agent follows an Observe-Think-Act loop to iteratively assemble and refine hybrid cardiac EP models. Right: the structured action space with a neural diffusion catalog and a physics-based reaction catalog. LEADS Overview. Cardiac EP is a domain rich in prior knowledge. The reaction-diffusion decomposition is physically well-established, and multiple reaction models have be… view at source ↗

**Figure 2.** Figure 2: Example LEADS agent trajectory on AP synthetic data. The agent starts with simple models, switches to more expressive architectures, diagnoses overfitting and applies regularization, and terminates after convergence. At each step, the LLM agent makes two decoupled decisions: selecting a reaction model from the catalog, and choosing one of four operations on the diffusion module: – Select: choose a new arch… view at source ↗

**Figure 3.** Figure 3: Activation time (AT) maps on synthetic data. We show examples on AP, RM and MS generated data respectively. Colors encode AT from early (blue) to late (red). LEADS closely recovers the ground-truth patterns without true reaction knowledge. While HDTwinGen has demonstrated strong performance on simpler dynamical systems [8], cardiac EP poses unique challenges—stable ODE integration on meshes, multi-variabl… view at source ↗

**Figure 4.** Figure 4: Activation time maps on real data. Each column shows a different method. Colors encode AT from early (blue) to late (red). LEADS can produce activation patterns closer to the ground truth than human-designed hybrid model. 4.4 Ablation Analysis We ablate key design choices of LEADS on all three synthetic datasets (Table. 3). Fixing the reaction to the ground-truth (Diff-Only) performs [PITH_FULL_IMAGE:figu… view at source ↗

read the original abstract

Building personalized cardiac electrophysiology (EP) digital twins requires identifying the appropriate model structure for each patient, not merely fitting parameters. Traditional methods rely on experts to manually prescribe hybrid physics-neural architectures, which requires deep domain expertise and does not transfer across patients. Recent works have applied large language models (LLMs) to generate or act as hybrid models. However, despite their promising generalization capacity, these LLM-based methods lack the structural priors needed for stable cardiac simulations. Hence, we propose LEADS, a framework that formulates cardiac EP domain knowledge as a structured action space and utilizes an LLM agent to discover hybrid models. The agent follows an iterative reasoning-and-action loop to select, combine, and refine hybrid models, whilst gradient descent handles parameter fitting. The proposed LEADS designs every candidate model towards physically grounded, interpretable, and numerically stable, while allowing open-ended architectural discovery. We validate LEADS on synthetic data with three ground-truth reaction models and on real cardiac EP data, demonstrating that it outperforms both human-designed hybrid models and other LLM-based hybrid modeling.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

LEADS puts an LLM agent inside a cardiac-EP-derived action space for iterative hybrid-model search; the abstract makes the outperformance claim but leaves the open-endedness and stability arguments hard to verify.

read the letter

The core idea is an LLM agent that iterates over a structured action space built from cardiac EP knowledge to pick, combine, and tweak hybrid reaction-diffusion components, then fits parameters by gradient descent. The paper positions this as a middle path between fully manual expert design and unconstrained LLM generation, and it reports better results than both on synthetic ground-truth cases and real EP recordings.

What stands out is the explicit encoding of domain priors into the agent's action vocabulary so that every candidate stays interpretable and numerically stable by construction. That is a concrete engineering choice worth noting for anyone trying to automate model structure search in physics-informed settings.

The soft spot is the one the stress-test flags. Because the action space is derived from existing domain knowledge, any model the agent produces is a recombination or refinement of components already known to experts. The abstract gives no concrete examples of structures that fall outside conventional forms, nor does it show how the space permits genuine architectural novelty rather than efficient enumeration inside a human-curated prior. Without the actual action-space definition or the discovered models, it is difficult to tell whether the reported gains come from better search or from the priors doing most of the work.

The validation is described only at a high level, so questions about exclusion criteria, error bars, and whether the synthetic cases are representative remain open. The work is still a serious attempt to close the loop between LLM reasoning and domain constraints.

This is mainly for groups already building patient-specific cardiac simulators who want ideas for reducing manual architecture tuning. It is coherent enough on its own terms to deserve a full referee read so the implementation details and quantitative results can be checked.

Referee Report

1 major / 0 minor

Summary. The manuscript proposes LEADS, a framework that encodes cardiac electrophysiology domain knowledge as a structured action space and employs an LLM agent in an iterative reasoning-and-action loop to discover hybrid physics-neural models for patient-specific cardiac EP digital twins. Gradient descent is used for parameter fitting within each candidate structure. The paper validates the approach on synthetic data generated from three ground-truth reaction models and on real cardiac EP recordings, claiming that the discovered models outperform both expert-designed hybrid architectures and prior LLM-based hybrid modeling methods while remaining physically grounded, interpretable, and numerically stable.

Significance. If the reported outperformance is reproducible and the discovered structures are shown to generalize beyond the action-space priors, the work would offer a practical route to automated model-structure identification for cardiac digital twins, lowering the barrier of deep domain expertise. The combination of agentic search with domain-derived constraints is a clear methodological contribution, though its impact depends on demonstrating that the discovered hybrids are not merely efficient recombinations of the enumerated components.

major comments (1)

[Abstract] Abstract: the central claim of 'open-ended architectural discovery' is in tension with the use of a 'structured action space' derived from cardiac EP domain knowledge. If the action space enumerates specific ionic-current terms, diffusion operators, or neural-replacement primitives, any discovered model is necessarily a combination or refinement of those primitives; the abstract supplies no evidence that structures outside conventional reaction-diffusion forms are reachable, which directly affects whether outperformance reflects genuine invention or improved search within expert priors.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their detailed review and for identifying a potential source of ambiguity in the abstract. We address the comment below with a point-by-point response and indicate where revisions will be made.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim of 'open-ended architectural discovery' is in tension with the use of a 'structured action space' derived from cardiac EP domain knowledge. If the action space enumerates specific ionic-current terms, diffusion operators, or neural-replacement primitives, any discovered model is necessarily a combination or refinement of those primitives; the abstract supplies no evidence that structures outside conventional reaction-diffusion forms are reachable, which directly affects whether outperformance reflects genuine invention or improved search within expert priors.

Authors: We agree that the abstract phrasing could be clarified to avoid any implication of discovery outside the reaction-diffusion paradigm. The structured action space is deliberately constructed from established cardiac EP primitives (ionic currents, diffusion operators, and neural-replacement modules) precisely to guarantee physical consistency, interpretability, and numerical stability—requirements that unconstrained LLM generation has historically failed to meet. Within this space the LLM agent performs open-ended combinatorial search: it iteratively selects, composes, and refines arbitrary subsets and parameterizations of the primitives, producing hybrid architectures that are not among the small number of hand-designed templates used in prior work. This is the sense in which we describe the discovery as open-ended: the search is not limited to a fixed catalogue of expert hybrids, yet every candidate remains inside the conventional reaction-diffusion family. The manuscript does not claim, nor does the experimental evidence demonstrate, the ability to invent new physical mechanisms beyond reaction-diffusion forms. To resolve the tension noted by the referee we will revise the abstract to read: “allowing open-ended architectural discovery of hybrid models within the reaction-diffusion framework.” This change will be incorporated in the revised manuscript. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation relies on external validation rather than internal reductions

full rationale

The provided abstract and description contain no equations, parameter-fitting procedures, or self-citations. The LEADS framework is presented as formulating domain knowledge into an action space for LLM-driven model discovery, followed by gradient-descent fitting and empirical validation on synthetic ground-truth models plus real cardiac data. No load-bearing step reduces a claimed prediction or uniqueness result to a fitted input, self-citation chain, or definitional equivalence. The central performance claims are therefore assessed against external benchmarks and remain self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Abstract-only review yields minimal concrete ledger entries; the primary unexamined premise is that domain knowledge can be reliably turned into an action space that guides stable discovery.

axioms (1)

domain assumption Cardiac EP domain knowledge can be effectively encoded as a structured action space for LLM agents
This encoding is presented as the enabling step for the agentic discovery loop.

invented entities (1)

LEADS framework no independent evidence
purpose: Agentic discovery of hybrid physics-neural models via structured action space
Newly introduced method whose stability and superiority are claimed but not inspectable from abstract.

pith-pipeline@v0.9.1-grok · 5728 in / 1246 out tokens · 29487 ms · 2026-06-27T00:46:27.980637+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

22 extracted references · 7 canonical work pages · 2 internal anchors

[1]

Chaos, Solitons & Fractals7(3), 293–301 (1996)

Aliev, R.R., Panfilov, A.V.: A simple two-variable model of cardiac excitation. Chaos, Solitons & Fractals7(3), 293–301 (1996)

1996
[2]

In: Forty-second International Conference on Machine Learning (2025)

Amad, H., Astorga, N., van der Schaar, M.: Continuously updating digital twins using large language models. In: Forty-second International Conference on Machine Learning (2025)

2025
[3]

Nature communications7(1), 11437 (2016)

Arevalo, H.J., Vadakkumpadan, F., Guallar, E., Jebb, A., Malamas, P., Wu, K.C., Trayanova, N.A.: Arrhythmia risk stratification of patients after myocardial infarc- tion using personalized heart models. Nature communications7(1), 11437 (2016)

2016
[4]

Computers in bi- ology and medicine134, 104476 (2021)

Bergquist, J.A., Good, W.W., Zenger, B., Tate, J.D., Rupp, L.C., MacLeod, R.S.: The electrocardiographic forward problem: A benchmark study. Computers in bi- ology and medicine134, 104476 (2021)

2021
[5]

Progress in biophysics and molecular biology104(1-3), 22–48 (2011)

Clayton, R.H., Bernus, O., Cherry, E., Dierckx, H., Fenton, F.H., Mirabella, L., Panfilov, A.V., Sachse, F.B., Seemann, G., Zhang, H.: Models of cardiac tissue electrophysiology: progress, challenges and open questions. Progress in biophysics and molecular biology104(1-3), 22–48 (2011)

2011
[6]

Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

Comanici, G., Bieber, E., Schaekermann, M., Pasupat, I., Sachdeva, N., Dhillon, I., Blistein, M., Ram, O., Zhang, D., Rosen, E., et al.: Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261 (2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025
[7]

arXiv preprint arXiv:2510.25536 (2025)

Du, B., Guo, M., He, S., Ye, Z., Zhu, X., Su, W., Zhu, S., Zhou, Y., Zhang, Y., Ai, Q., et al.: Twinvoice: A multi-dimensional benchmark towards digital twins via llm persona simulation. arXiv preprint arXiv:2510.25536 (2025)

work page arXiv 2025
[8]

Advances in Neural Information Processing Systems37, 72170–72218 (2024)

Holt, S., Liu, T., van der Schaar, M.: Automatically learning hybrid digital twins of dynamical systems. Advances in Neural Information Processing Systems37, 72170–72218 (2024)

2024
[9]

arXiv preprint arXiv:2403.15433 (2024)

Jiang, X., Vadhavkar, S., Ye, Y., Toloubidokhti, M., Missel, R., Wang, L.: Hyper- ep: Meta-learning hybrid personalized models for cardiac electrophysiology. arXiv preprint arXiv:2403.15433 (2024)

work page arXiv 2024
[10]

In: Proceedings of the 10th ACM International Conference on Systems for Energy- Efficient Buildings, Cities, and Transportation

Li, M., Wang, R., Zhou, X., Zhu, Z., Wen, Y., Tan, R.: Chattwin: Toward au- tomated digital twin generation for data center via large language models. In: Proceedings of the 10th ACM International Conference on Systems for Energy- Efficient Buildings, Cities, and Transportation. pp. 208–211 (2023)

2023
[11]

arXiv preprint arXiv:2502.14642 (2025)

Li, R., Xia, H., Yuan, X., Dong, Q., Sha, L., Li, W., Sui, Z.: How far are llms from being our digital twins? a benchmark for persona-based behavior chain simulation. arXiv preprint arXiv:2502.14642 (2025)

work page arXiv 2025
[12]

Bulletin of mathematical biology65(5), 767–793 (2003)

Mitchell, C.C., Schaeffer, D.G.: A two-current model for the dynamics of cardiac membrane. Bulletin of mathematical biology65(5), 767–793 (2003)

2003
[13]

Frontiers in physiology8, 668 (2017) 10 Z

Passini, E., Britton, O.J., Lu, H.R., Rohrbacher, J., Hermans, A.N., Gallacher, D.J., Greig, R.J., Bueno-Orovio, A., Rodriguez, B.: Human in silico drug tri- als demonstrate higher accuracy than animal models in predicting clinical pro- arrhythmic cardiotoxicity. Frontiers in physiology8, 668 (2017) 10 Z. Zhou et al

2017
[14]

Nature biomedical engineering2(10), 732–740 (2018)

Prakosa, A., Arevalo, H.J., Deng, D., Boyle, P.M., Nikolov, P.P., Ashikaga, H., Blauer, J.J., Ghafoori, E., Park, C.J., Blake III, R.C., et al.: Personalized virtual- heart technology for guiding the ablation of infarct-related ventricular tachycardia. Nature biomedical engineering2(10), 732–740 (2018)

2018
[15]

Advances in Neural Information Processing Systems34, 11364–11383 (2021)

Qian, Z., Zame, W., Fleuren, L., Elbers, P., van der Schaar, M.: Integrating expert odes into neural odes: pharmacology and disease progression. Advances in Neural Information Processing Systems34, 11364–11383 (2021)

2021
[16]

IEEE Transactions on Biomedical Engineering 41(8), 743–757 (2002)

Rogers, J.M., McCulloch, A.D.: A collocation-galerkin finite element model of car- diac action potential propagation. IEEE Transactions on Biomedical Engineering 41(8), 743–757 (2002)

2002
[17]

From mind to machine: The rise of manus ai as a fully autonomous digital agent.arXiv preprint arXiv:2505.02024, 2025

Shen, M., Li, Y., Chen, L., Yang, Q.: From mind to machine: The rise of manus ai as a fully autonomous digital agent. arXiv preprint arXiv:2505.02024 (2025)

work page arXiv 2025
[18]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Toloubidokhti, M., Missel, R., Lian, S., Wang, L.: Meta-learning physics-informed neural networks for personalized cardiac modeling. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 344–354. Springer (2025)

2025
[19]

Graph Attention Networks

Veličković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph attention networks. arXiv preprint arXiv:1710.10903 (2017)

work page internal anchor Pith review Pith/arXiv arXiv 2017
[20]

Wehenkel, A., Behrmann, J., Hsu, H., Sapiro, G., Louppe, G., Jacobsen, J.H.: Robusthybridlearningwithexpertaugmentation.arXivpreprintarXiv:2202.03881 (2022)

work page arXiv 2022
[21]

Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K.R., Cao, Y.: React: Synergizingreasoningandactinginlanguagemodels.In:Theeleventhinternational conference on learning representations (2022)

2022
[22]

Journal of Statistical Mechanics: Theory and Experiment2021(12), 124012 (2021)

Yin, Y., Le Guen, V., Dona, J., De Bézenac, E., Ayed, I., Thome, N., Gallinari, P.: Augmenting physical models with deep networks for complex dynamics forecasting. Journal of Statistical Mechanics: Theory and Experiment2021(12), 124012 (2021)

2021

[1] [1]

Chaos, Solitons & Fractals7(3), 293–301 (1996)

Aliev, R.R., Panfilov, A.V.: A simple two-variable model of cardiac excitation. Chaos, Solitons & Fractals7(3), 293–301 (1996)

1996

[2] [2]

In: Forty-second International Conference on Machine Learning (2025)

Amad, H., Astorga, N., van der Schaar, M.: Continuously updating digital twins using large language models. In: Forty-second International Conference on Machine Learning (2025)

2025

[3] [3]

Nature communications7(1), 11437 (2016)

Arevalo, H.J., Vadakkumpadan, F., Guallar, E., Jebb, A., Malamas, P., Wu, K.C., Trayanova, N.A.: Arrhythmia risk stratification of patients after myocardial infarc- tion using personalized heart models. Nature communications7(1), 11437 (2016)

2016

[4] [4]

Computers in bi- ology and medicine134, 104476 (2021)

Bergquist, J.A., Good, W.W., Zenger, B., Tate, J.D., Rupp, L.C., MacLeod, R.S.: The electrocardiographic forward problem: A benchmark study. Computers in bi- ology and medicine134, 104476 (2021)

2021

[5] [5]

Progress in biophysics and molecular biology104(1-3), 22–48 (2011)

Clayton, R.H., Bernus, O., Cherry, E., Dierckx, H., Fenton, F.H., Mirabella, L., Panfilov, A.V., Sachse, F.B., Seemann, G., Zhang, H.: Models of cardiac tissue electrophysiology: progress, challenges and open questions. Progress in biophysics and molecular biology104(1-3), 22–48 (2011)

2011

[6] [6]

Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

Comanici, G., Bieber, E., Schaekermann, M., Pasupat, I., Sachdeva, N., Dhillon, I., Blistein, M., Ram, O., Zhang, D., Rosen, E., et al.: Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261 (2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025

[7] [7]

arXiv preprint arXiv:2510.25536 (2025)

Du, B., Guo, M., He, S., Ye, Z., Zhu, X., Su, W., Zhu, S., Zhou, Y., Zhang, Y., Ai, Q., et al.: Twinvoice: A multi-dimensional benchmark towards digital twins via llm persona simulation. arXiv preprint arXiv:2510.25536 (2025)

work page arXiv 2025

[8] [8]

Advances in Neural Information Processing Systems37, 72170–72218 (2024)

Holt, S., Liu, T., van der Schaar, M.: Automatically learning hybrid digital twins of dynamical systems. Advances in Neural Information Processing Systems37, 72170–72218 (2024)

2024

[9] [9]

arXiv preprint arXiv:2403.15433 (2024)

Jiang, X., Vadhavkar, S., Ye, Y., Toloubidokhti, M., Missel, R., Wang, L.: Hyper- ep: Meta-learning hybrid personalized models for cardiac electrophysiology. arXiv preprint arXiv:2403.15433 (2024)

work page arXiv 2024

[10] [10]

In: Proceedings of the 10th ACM International Conference on Systems for Energy- Efficient Buildings, Cities, and Transportation

Li, M., Wang, R., Zhou, X., Zhu, Z., Wen, Y., Tan, R.: Chattwin: Toward au- tomated digital twin generation for data center via large language models. In: Proceedings of the 10th ACM International Conference on Systems for Energy- Efficient Buildings, Cities, and Transportation. pp. 208–211 (2023)

2023

[11] [11]

arXiv preprint arXiv:2502.14642 (2025)

Li, R., Xia, H., Yuan, X., Dong, Q., Sha, L., Li, W., Sui, Z.: How far are llms from being our digital twins? a benchmark for persona-based behavior chain simulation. arXiv preprint arXiv:2502.14642 (2025)

work page arXiv 2025

[12] [12]

Bulletin of mathematical biology65(5), 767–793 (2003)

Mitchell, C.C., Schaeffer, D.G.: A two-current model for the dynamics of cardiac membrane. Bulletin of mathematical biology65(5), 767–793 (2003)

2003

[13] [13]

Frontiers in physiology8, 668 (2017) 10 Z

Passini, E., Britton, O.J., Lu, H.R., Rohrbacher, J., Hermans, A.N., Gallacher, D.J., Greig, R.J., Bueno-Orovio, A., Rodriguez, B.: Human in silico drug tri- als demonstrate higher accuracy than animal models in predicting clinical pro- arrhythmic cardiotoxicity. Frontiers in physiology8, 668 (2017) 10 Z. Zhou et al

2017

[14] [14]

Nature biomedical engineering2(10), 732–740 (2018)

Prakosa, A., Arevalo, H.J., Deng, D., Boyle, P.M., Nikolov, P.P., Ashikaga, H., Blauer, J.J., Ghafoori, E., Park, C.J., Blake III, R.C., et al.: Personalized virtual- heart technology for guiding the ablation of infarct-related ventricular tachycardia. Nature biomedical engineering2(10), 732–740 (2018)

2018

[15] [15]

Advances in Neural Information Processing Systems34, 11364–11383 (2021)

Qian, Z., Zame, W., Fleuren, L., Elbers, P., van der Schaar, M.: Integrating expert odes into neural odes: pharmacology and disease progression. Advances in Neural Information Processing Systems34, 11364–11383 (2021)

2021

[16] [16]

IEEE Transactions on Biomedical Engineering 41(8), 743–757 (2002)

Rogers, J.M., McCulloch, A.D.: A collocation-galerkin finite element model of car- diac action potential propagation. IEEE Transactions on Biomedical Engineering 41(8), 743–757 (2002)

2002

[17] [17]

From mind to machine: The rise of manus ai as a fully autonomous digital agent.arXiv preprint arXiv:2505.02024, 2025

Shen, M., Li, Y., Chen, L., Yang, Q.: From mind to machine: The rise of manus ai as a fully autonomous digital agent. arXiv preprint arXiv:2505.02024 (2025)

work page arXiv 2025

[18] [18]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Toloubidokhti, M., Missel, R., Lian, S., Wang, L.: Meta-learning physics-informed neural networks for personalized cardiac modeling. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 344–354. Springer (2025)

2025

[19] [19]

Graph Attention Networks

Veličković, P., Cucurull, G., Casanova, A., Romero, A., Lio, P., Bengio, Y.: Graph attention networks. arXiv preprint arXiv:1710.10903 (2017)

work page internal anchor Pith review Pith/arXiv arXiv 2017

[20] [20]

Wehenkel, A., Behrmann, J., Hsu, H., Sapiro, G., Louppe, G., Jacobsen, J.H.: Robusthybridlearningwithexpertaugmentation.arXivpreprintarXiv:2202.03881 (2022)

work page arXiv 2022

[21] [21]

Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K.R., Cao, Y.: React: Synergizingreasoningandactinginlanguagemodels.In:Theeleventhinternational conference on learning representations (2022)

2022

[22] [22]

Journal of Statistical Mechanics: Theory and Experiment2021(12), 124012 (2021)

Yin, Y., Le Guen, V., Dona, J., De Bézenac, E., Ayed, I., Thome, N., Gallinari, P.: Augmenting physical models with deep networks for complex dynamics forecasting. Journal of Statistical Mechanics: Theory and Experiment2021(12), 124012 (2021)

2021