Research map / Probabilistic and scientific modeling / Differential equations and PDEs
Med-Real2Sim: Non-Invasive Medical Digital Twins using Physics-Informed Self-Supervised Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Med-Real2Sim pretrains a differentiable neural surrogate of a lumped-parameter cardiac circuit model and then freezes it while finetuning an echocardiography 3D-CNN to infer digital-twin parameters, enabling unsupervised disease detection and in-silico LVAD trials.
Problem
Identifying patient-specific physics-based digital twin parameters (e.g., cardiac pressure-volume dynamics) from noninvasive data when paired parameter-measurement datasets are unavailable, so standard supervised learning cannot be applied.
Approach
Physics-informed self-supervised learning: pretrain a feedforward network to emulate the forward physics model on synthetic parameter-state pairs, freeze it, and finetune an image encoder so that its predicted parameters pass through the frozen emulator to match observed noninvasive labels (LV volumes at ES and ED).
Key results
- Med-Real2Sim achieves EF MAE of 6.81% on CAMUS and 5.40% on EchoNet, comparable to a supervised 3DCNN baseline (6.58% and 5.62%).
- Average EF increases by 17.6% (CAMUS) and 18.9% (EchoNet) in silico after adding a simulated LVAD.
- Simulated PV-loop features separate mitral stenosis patients, a condition not used in training, while EF alone has AUC of 0.5 for MS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Med-Real2Sim instead of supervised 3DCNN | EF prediction · CAMUS test | MAE (EF, %): 6.58 -> 6.81 | Hurt |
| use Med-Real2Sim instead of supervised 3DCNN | EF prediction · EchoNet test | MAE (EF, %): 5.62 -> 5.40 | Helped |
| use the physics-informed forward model rather than a Neural ODE | Cardiac hemodynamics forward modeling · Cardiac hemodynamics | MAE (EF): 3.39 (N = 20) -> 2.30 (N=1000) | Helped |
| use the physics-informed forward model rather than a PINN | Cardiac hemodynamics forward modeling · Cardiac hemodynamics | MAE (EF): 7.67 (N=20) -> 2.30 (N=1000) | Helped |
| add LVAD intervention | Med-Real2Sim cardiac digital twins · CAMUS population | EF change: 17.6% | Helped |
| add LVAD intervention | Med-Real2Sim cardiac digital twins · EchoNet population | EF change: 18.9% | Helped |
Surprising findings
- EF alone has AUC of about 0.5 for predicting mitral stenosis, while the digital twin trend captures MS-related patterns.
Limitations
- No clinical validation; cannot compare to invasive resistance or LV pressure measurements.
- Relies on publicly available datasets with limited labels (only volume labels).
- Model fidelity bounded by the fixed lumped-parameter circuit model and literature-fixed parameters.
Experimental setup
Datasets: CAMUS EchoNet
Benchmarks: CAMUS EchoNet
Metrics: MAE on ejection fraction (%) AUC
Baselines: Supervised 3DCNN PINN Neural ODE
Builds on: physics-informed neural networks Neural ODE Windkessel lumped-parameter model
Closest papers by meaning
- Continuously Updating Digital Twins using Large Language Models (ICML 2025 · similarity 0.83)
- Continual Self-supervised Learning: Towards Universal Multi-modal Medical Data Representation Learning (CVPR 2024 · similarity 0.83)
- Integrating Sequence and Image Modeling in Irregular Medical Time Series Through Self-Supervised Learning (AAAI 2025 · similarity 0.83)
- DrFuse: Learning Disentangled Representation for Clinical Multi-Modal Fusion with Missing Modality and Modal Inconsistency (AAAI 2024 · similarity 0.83)
- CLIMB: Data Foundations for Large Scale Multimodal Clinical Foundation Models (ICML 2025 · similarity 0.83)
- From Token to Rhythm: A Multi-Scale Approach for ECG-Language Pretraining (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.