⚡ electrical engineering

Graph-Based ECG Synthesis with Activation-Consistency Certification and Diagnostics-Aware Morphology Curation

This paper presents a graph-based ECG synthesis framework that integrates activation-consistency certification via Bellman residuals with diagnostics-aware morphology curation to generate high-fidelity, controllable synthetic signals that outperform traditional eikonal-template methods in both activation accuracy and morphological diversity.

Original authors: Yongjun Yoo

Published 2026-06-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yongjun Yoo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to recognize heartbeats. You need thousands of practice heartbeats (ECGs) to train it, but you can't just use real patient data because of privacy laws, and you can't just make up random squiggly lines because they might look like heartbeats but behave like nonsense.

This paper presents a new "factory" for making fake but scientifically accurate heartbeats. The goal is to create synthetic heart signals that look real enough to train AI, but are built on a solid mathematical foundation so we know exactly how they were made.

Here is how their system works, broken down into simple concepts:

1. The Blueprint: A "Heart Map" (The Graph)

Instead of building a complex, heavy 3D model of a heart (which is slow and hard to tweak), the researchers built a simplified map of the heart. Think of this like a subway map:

Stations are points in the heart muscle.
Tracks are the pathways electricity travels.
Speed limits are set for how fast electricity moves on different tracks.

This map is the "skeleton" that both of their methods use. It keeps things fast and reproducible.

2. The Two Factory Lines: ET and RE

The researchers built two different machines (backends) to generate the heartbeat signals using this same map.

Machine A (ET - The "Template" Machine): This is the fast, simple version. It calculates when the electricity hits each station on the map, then simply pastes a pre-made "heartbeat shape" onto that timing. It's like a stamping press: fast and consistent, but the shape of the heartbeat is fixed.
Machine B (RE - The "Reaction" Machine): This is the fancy version. It still uses the same timing map, but instead of just stamping a shape, it simulates the actual chemical "reaction" of the heart cells. It adds a little bit of "smoothing" (like a diffuser on a camera lens) to make the recovery phase of the heartbeat look more natural. It's slower, but it might produce more realistic shapes.

3. The Quality Control: The "Bellman Certificate"

How do you know the fake heartbeat isn't broken? If the electricity travels backward or gets stuck, the signal is garbage.

The researchers created a mathematical "seal of approval" called the Bellman Certificate.

The Analogy: Imagine a delivery driver trying to get from Point A to Point B. The "Eikonal" method calculates the fastest route. The "Bellman Certificate" is a quick math check that asks: "Did the driver actually take the fastest route, or did they take a weird detour?"
If the math check passes (the "residual" is low), the system knows the timing of the heartbeat is logically consistent. If it fails, the heartbeat is thrown out before it even gets to the next step.

4. The "Doctor's Filter" (Diagnostics-Aware Curation)

Even if the timing is perfect, the heartbeat might still look weird (e.g., the T-wave, which represents the heart resetting, might be upside down).

The researchers built a two-step filter:

Step 1 (The Calculator): It measures specific things: Is the P-wave (the first bump) the right height? Is the QRS complex (the big spike) the right width?
Step 2 (The Policy): This is where the "rules" come in. Depending on what the experiment needs, the system decides: "Okay, this heartbeat is mathematically correct, but for our specific goal, we need the T-wave to be positive in Lead II. If it's negative, reject it."

This separates the measurement (what the signal looks like) from the decision (whether we keep it).

5. The Results: What Did They Find?

Consistency: The fancy "Reaction" machine (RE) stayed very close to the simple "Template" machine (ET) regarding timing. The math certificate confirmed the timing was nearly perfect (within a millisecond).
Better Shapes: The fancy machine (RE) produced slightly more "acceptable" heartbeats than the simple one. Out of 2,000 attempts, the RE machine passed the final quality filter 658 times, while the ET machine only passed 578 times.
Coverage: The RE machine also covered a slightly wider variety of "healthy" heartbeat shapes, meaning it's better at generating diverse data for training.
Real-World Test: They tried using these fake heartbeats to help train an AI to spot real heartbeats with very little real data. The fake data helped the AI learn slightly faster and more accurately, especially for the smaller parts of the heartbeat (the P-wave).

The Bottom Line

This paper doesn't claim to have created a perfect replacement for real patient data. Instead, it offers a controlled, transparent, and mathematically verified way to generate synthetic heartbeats.

It's like building a high-quality "training dummy" for a robot. The researchers proved that their dummy moves in a way that is physically consistent (thanks to the Bellman Certificate) and looks realistic enough to be useful for training AI, especially when they use the more advanced "Reaction" machine.

Technical Summary: Graph-Based ECG Synthesis with Activation-Consistency Certification and Diagnostics-Aware Morphology Curation

Problem Statement
Synthetic electrocardiogram (ECG) generation is valuable for algorithm development, robustness evaluation, and data augmentation, particularly when real recordings are limited by privacy, annotation costs, or class imbalance. However, a critical trade-off exists between realism, controllability, and quality control. Data-driven generative models often lack interpretable connections to physiological mechanisms, while computational electrophysiology models can produce numerically valid but morphologically implausible signals. Existing methods frequently lack standardized quality-assessment criteria, specifically failing to ensure that synthetic signals preserve interpretable activation order, P-QRS-T structure, and lead-wise consistency. This study addresses the need for a framework that combines mechanistic interpretability with rigorous, computable quality control.

Methodology
The authors propose a graph-based ECG synthesis framework built on a unified heart graph $G=(V, E)$ representing dual-layer atrial and ventricular structures, a specialized conduction system, and fibrous insulation. The framework operates through three primary layers:

Activation-Consistency Certification:
- The graph Eikonal activation problem is formulated as a Bellman fixed-point problem.
- A Bellman residual is derived as a computable certificate to verify activation-time consistency. If a candidate activation field has a small residual, it is uniformly close to the true Eikonal time (Theorem 2).
- To handle global time-scale mismatches, an affine calibration ( $\tau \approx \alpha T + \beta$ ) is applied, estimating parameters via least-squares regression.
Unified Synthesis Pipeline with Dual Backends:
- A shared forward pipeline projects transmembrane potentials ( $V_m$ ) to 12-lead ECGs via extracellular potentials and torso potentials.
- Two solver backends operate on the same graph:
  - ET (Eikonal-Template): A fast baseline where $V_m$ is a prescribed action-potential template shifted by the Eikonal activation time. Recovery is fixed by the template.
  - RE (Reaction-Eikonal): Retains the Eikonal activation backbone but replaces the template shift with triggered ionic dynamics (Aliev–Panfilov or Bueno-Orovio models) and a pseudo-diffusion term. This term couples neighboring nodes via a graph Laplacian with strength $\kappa$ , smoothing recovery gradients without distorting the activation backbone.
Diagnostics-Aware Curation Pipeline:
- A two-stage evaluation process separates metric computation from experiment-specific acceptance policies.
- Stage 1 (Metric Computation): Computes activation-field diagnostics (e.g., sequence reversal checks, causal predecessor consistency), ECG morphology/interval metrics (anchored by latent fields), recovery plausibility scores ( $s_{rep}$ ), and feature-space coverage.
- Stage 2 (Policy-Based Acceptance): Applies hard filters (e.g., gross activation failures, missing QRS) and soft constraints (e.g., T-wave polarity, recovery scores) tailored to specific experimental goals (throughput screening vs. final morphology curation).

Key Contributions

Bellman Certificate: Formulation of graph Eikonal activation as a Bellman fixed-point problem and derivation of a residual-based certificate for activation-time consistency.
Unified Pipeline: Development of a graph-based synthesis framework supporting both ET and RE backends with separated activation and recovery controls and a shared 12-lead forward model.
Curation Framework: Introduction of a diagnostics-aware pipeline that decouples metric computation from acceptance policies, enabling diverse analyses from throughput screening to balanced multi-lead morphology curation.
Validation: Empirical validation showing that the RE backend preserves activation controllability while offering modest improvements in morphology yield and coverage compared to ET.

Results

Certificate Validation: On the cardiac graph (1321 nodes), activation times extracted from the RE backend showed near-millisecond agreement with the Eikonal backbone after causal predecessor filtering ( $R^2 = 0.99876$ , Bellman residual $\approx 0.92$ ms). No predecessor cycles were detected across time-step variations.
Activation Controllability: In activation-atlas experiments, RE tracked ET closely in PR timing, QRS range, and amplitude responses to conductivity knobs (AV, annulus, atrial), confirming that adding pseudo-diffusion dynamics does not disrupt activation control.
Recovery Dynamics: Experiments demonstrated that the endo-epicardial APD gradient is the primary determinant of the admissible T-wave morphology window. The pseudo-diffusion strength $\kappa$ acted as a secondary smoothing parameter, improving local repolarization smoothness without significantly shifting QT timing.
Curation Outcomes: Under a balanced multi-lead curation policy (E6), the RE backend accepted 658/2000 samples versus 578/2000 for ET. Feature-space coverage increased from 0.09248 (ET) to 0.09888 (RE).
Downstream Utility: In a low-label P/QRS delineation task on PTB-XL+, synthetic pretraining provided small but consistent improvements in P-wave segmentation and boundary localization, particularly at low label fractions (1% and 10%), suggesting the signals provide useful initialization structure.

Significance and Claims
The paper claims to provide a conservative basis for controllable and curated synthetic ECG generation. The significance lies in treating activation consistency, morphology control, and sample curation as coupled but separable components.

The framework demonstrates that a Bellman residual can serve as a practical, computable consistency check for activation fields, ensuring the synthetic signals adhere to the underlying activation backbone.
The RE backend is shown to preserve the activation controllability of the ET baseline while adding recovery-aware morphology modulation.
The authors explicitly state that the improvement offered by RE is modest and does not establish broad superiority over ET; rather, it offers a trade-off where recovery-aware control is needed.
The study emphasizes that the generated ECGs are not diagnostic-grade clinical replacements but serve as mechanism-consistent initialization signals for downstream tasks. The framework prioritizes reproducibility and transparency over subject-specific complexity, avoiding heavy preprocessing stacks.