🤖 machine learning

Neural Surrogate Forward Modelling For Electrocardiology Without Explicit Intracellular Conductivity Tensor

This proof-of-concept study introduces a deep learning-based neural surrogate that accurately maps left atrial intracellular potentials to far-field ECGs without requiring explicit, unmeasurable intracellular conductivity tensors, thereby reducing structural uncertainty in non-invasive cardiac electrophysiology modeling.

Original authors: Shaheim Ogbomo-Harmitt, Cesare Magnetti, Jakub Grzelak, Oleg Aslanidi

Published 2026-05-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shaheim Ogbomo-Harmitt, Cesare Magnetti, Jakub Grzelak, Oleg Aslanidi

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 your heart is a complex city, and the electrical signals that make it beat are like traffic flowing through its streets. Doctors often want to predict what the "traffic report" looks like from the outside (the ECG on your skin) based on what's happening inside the city (the electrical activity in the heart muscle).

Traditionally, to make this prediction, scientists have to build a massive, complicated mathematical map of the city. They have to guess exactly how "smooth" or "slippery" the roads are (conductivity) and which way the traffic flows (fiber direction). The problem is, they can't actually measure these road conditions in a living patient. They have to guess, and if their guess is wrong, the traffic report they generate is also wrong.

The New Approach: Learning the Pattern Instead of Drawing the Map

This paper presents a clever new idea: instead of trying to build a perfect map of the roads, let's train a smart computer (an AI) to just "learn the pattern" of how the inside traffic turns into an outside report.

Think of it like learning to predict the weather.

The Old Way: You try to calculate every single molecule of air, humidity, and wind speed using complex physics equations. If you get one number wrong (like the wind speed), your forecast fails.
The New Way (This Paper): You show the computer thousands of photos of the sky and the resulting weather reports. The computer learns, "Oh, when the sky looks this specific way, the report usually says that." It doesn't need to know the physics of wind; it just needs to recognize the pattern.

How They Taught the Computer

The researchers created a "training school" for their AI using data from 74 different heart shapes (simulated on a computer).

The Input: They fed the AI the electrical voltage maps from the inside of the left atrium (the heart's upper chamber).
The Output: They showed the AI the resulting ECG signal (the line you see on a heart monitor).
The Magic Trick: The AI was designed with a special "encoder" (a translator). It didn't just look at the numbers; it looked at the shape of the heart and the flow of the electricity simultaneously. It used a technique called "spectral propagation," which is like listening to the echo of a sound in a room to understand the room's shape, even if you can't see the walls.

Crucially, the AI never saw the "road conditions" (the intracellular conductivity tensor) during the test. It learned to skip that step entirely. It figured out that if it looks at the electrical pattern closely enough, it can guess the ECG result without needing to know the specific properties of the heart tissue.

The Results

The AI was tested on new heart data it had never seen before.

Accuracy: It got a score of 0.949 out of 1.0 (where 1.0 is perfect). This means the AI's predicted ECG lines looked almost identical to the real ones.
The Surprise: It did this incredibly well even though it was only trained on 74 examples. Usually, AI needs thousands of examples to learn this well.

What This Means (According to the Paper)

The paper claims this is a "proof-of-concept." It proves that you don't need to guess the unknown properties of heart tissue to get a good ECG prediction. By letting the AI learn the direct link between the inside signal and the outside signal, they removed a major source of error (the guesswork about tissue properties).

The authors suggest that if they train this system on more people and more complex heart rhythms (like the chaotic traffic of Atrial Fibrillation), it could make non-invasive heart imaging much more accurate. But for now, the main takeaway is simply: We found a way to predict the heart's external signal without needing to know the invisible internal details we usually have to guess.

Technical Summary: Neural Surrogate Forward Modelling for Electrocardiology

Problem Statement
Accurate forward modelling is a prerequisite for non-invasive cardiac electrophysiology, particularly for the inverse problem of reconstructing cardiac activation from body surface potentials. This is critical in the context of atrial fibrillation (AF), where electrical activation is highly disorganised. Conventional physics-based forward models rely on the explicit specification of intracellular conductivity tensors, which include tissue conductivity, myocardial fibre direction, and anisotropy ratios. However, these parameters cannot be directly measured in routine clinical practice and must be assumed or estimated. This necessity introduces structural modelling errors that propagate into inverse solutions, limiting the accuracy of non-invasive reconstructions. While previous efforts have focused on refining torso electrical properties, the error introduced by assumed intracellular conductivity tensors remains a significant barrier.

Methodology
This study proposes a deep learning (DL) framework that learns a direct mapping from left atrial (LA) intracellular electrical potentials to far-field Lead II ECGs, bypassing the need for explicit intracellular conductivity inputs at inference time.

Data Generation: Atrial electrophysiology was simulated using the openCARP platform with a monodomain model and an AF-adapted Courtemanche–Ramírez–Nattel cell formulation. Simulations utilized 93 patient-specific LA surface meshes (74 used for training/validation/testing after excluding 7 for instability) with fibre orientations mapped from an atlas. While conduction velocity (0.7 ms⁻¹) and anisotropy ratios (4:1) were fixed, subject-specific fibre orientations generated spatially varying, geometry-dependent conductivity tensor fields. Lead II ECGs were computed via the infinite-volume conductor method and filtered (0.05–60 Hz) to match clinical standards.
Encoder Architecture: The model employs a sequence-to-sequence architecture. The encoder maps 3D coordinates and transmembrane voltages on the atrial surface to a fixed-dimensional latent vector through three distinct feature streams:
1. Local Geometric Encoding: Utilizes DiffusionNet with precomputed geometric operators (cotangent Laplacian, mass matrix, intrinsic gradients, and Laplace–Beltrami eigenpairs) to extract local embeddings robust to mesh density changes.
2. Global Spectral Propagation: Applies a learnable spectral response function, $g(\lambda) = e^{-s\lambda} + b$ , to the local features using the first 64 Laplace–Beltrami eigenfunctions. This allows for efficient long-range, geometry-aware coupling without the oversmoothing associated with stacked diffusion blocks.
3. Raw Voltage Projection: A lightweight MLP maps raw voltage values to feature vectors to preserve sharp voltage transitions.
  These streams are fused via a per-node gating mechanism (softmax-normalised coefficients) and summarised using a quadrature procedure guided by Universal Atrial Coordinates (UAC), incorporating mass-weighted pooling and global voltage statistics.
Decoder Architecture: The decoder maps the latent sequence to the ECG trace using sinusoidal time embeddings, time-aware Bahdanau attention, and an LSTM network. It computes latent differences ( $\Delta z_t$ ) to capture rapid temporal transitions and uses an autoregressive procedure to generate the full waveform.
Training: The model was trained on 74 subjects (80/10/10 split) using a hybrid loss function combining a Huber time-domain loss and a spectral-entropy loss (weighted by a cosine decay schedule) to penalise differences in power spectral densities.

Key Results
The proposed framework achieved a coefficient of determination ( $R^2$ ) of 0.949 ± 0.037 and a mean absolute error (MAE) of 0.0036 ± 0.0015 mV on the test set. The model successfully predicted ECG morphology and timing with high fidelity, as demonstrated in representative examples where predicted signals closely matched ground-truth waveforms. An ablation study confirmed that removing specific encoder components led to decreased performance.

Significance and Claims
This work presents a proof-of-concept for a data-driven forward solution in electrocardiology that eliminates the need for explicit intracellular conductivity specification during inference. By learning to account for anisotropic conduction effects implicitly through subject-specific fibre orientations in the training data, the model reduces the structural uncertainty associated with assumed conductivity tensors. The authors claim that this approach has the potential to improve the accuracy of inverse atrial electrophysiology solutions, particularly in the challenging context of AF. The study suggests that while the current model was trained on a relatively small cohort (74 subjects), incorporating larger and more diverse datasets could further enhance generalisability. Future directions identified by the authors include extending the framework to whole-heart or biatrial geometries, predicting multiple ECG leads simultaneously, and modelling ECGs associated with reentrant electrical activity.

Technical Summary: Neural Surrogate Forward Modelling for Electrocardiology

More like this