Skip to main content

m39 · AI on intracardiac electrograms and maps

Warning

Education only. Not for diagnosis, treatment or any clinical decision.

In 60 seconds

Electrograms recorded inside the heart carry more detail than the surface ECG, and mapping systems collect thousands of them, each tied to a place on the heart. AI is used on both: on single electrograms or short recordings, to label patterns such as rotational activation (Alhusseini et al. 2020) or driver sites (Zolotarev et al. 2020); and on whole maps, where points are linked by the mesh and a graph network can learn from neighbours (Saluja et al. 2024). One AI tool has been tested in a randomized trial: ablation guided by an algorithm that detects spatiotemporal electrogram dispersion, added to pulmonary vein isolation (PVI), against PVI alone (Deisenhofer, Albenque, et al. 2025). This module derives graph convolution on a mapping mesh, the metrics for judging a predicted region, and losses for rare targets, and it shows how a split by window rather than by patient overstates accuracy on real intracardiac recordings.

From the field

  • Experts label, then networks learn. A Stanford team labeled 175,000 bi-atrial activation grids from 35 patients for rotational activation (expert agreement κ = 0.79) and trained a convolutional network that reached 95.0% accuracy on a separate test set of 50,000 grids (Alhusseini et al. 2020). Labels from experts carry their disagreement into the model (m35).
  • Ground truth from another modality. In explanted human atria, optical mapping defined which electrodes sat on a driver. Averaging features over each electrode’s 8 neighbours improved classification over single electrograms (Zolotarev et al. 2020), an early form of what graph networks do.
  • Graphs fit mapping data. A graph convolutional network on electroanatomic maps of 29 scar-related atrial tachycardias predicted isthmus areas with a median centroid distance of 7.3 mm in the test group, but a mean Dice overlap of only 13.9% (Saluja et al. 2024). The two metrics answer different questions (demo m39-d2).
  • Rare targets need different metrics. A 1-D network for ganglionated plexi on bipolar electrograms, with target windows near 3.5% of the data and trained with a focal loss, had a recall of 0.85 but a precision of 0.09 on validation (Guler et al. 2026).
  • Signals and outcomes. Stanford groups have combined intracardiac electrograms, the ECG and clinical data to predict AF recurrence after ablation (Tang et al. 2022), and found that acute AF termination and 1-year success were predicted by different features (Ganesan et al. 2025).
  • Unsupervised structure. Gaussian mixtures fitted to the whole distribution of left atrial voltages split one cohort into milder and more severe atrial cardiomyopathy without a fixed threshold (Huttelmaier et al. 2026).

Derivations

Symbols: a mesh with \(n\) vertices, adjacency \(A\) (\(A_{ij} = 1\) when vertices \(i\) and \(j\) share an edge), degree \(d_i\), feature matrix \(X\) (\(n \times f\)), weights \(W\), \(p\) a predicted probability, \(y\) a label, \(\alpha\) and \(\gamma\) the focal-loss parameters.

Graph convolution on a mapping mesh

A mapping system’s points live on a triangle mesh. A graph convolution layer updates every vertex from itself and its neighbours (Kipf and Welling 2017):

\[ H' = \sigma\!\left(\hat A H W\right), \qquad \hat A = \tilde D^{-1/2}\,(A + I)\,\tilde D^{-1/2} \tag{1}\]

where \(\tilde D\) is the diagonal matrix of degrees of \(A + I\) and \(\sigma\) a nonlinearity such as ReLU. Each vertex takes a weighted average of its own features and its neighbours’, \(\hat A_{ij} = 1/\sqrt{\tilde d_i \tilde d_j}\), then mixes the feature channels with \(W\), which is shared by all vertices. After \(L\) layers a vertex depends only on vertices within \(L\) edges: the graph version of the receptive field in m37.

Without \(W\) and \(\sigma\), \(\hat A\) is a smoothing operator. It is symmetric, its eigenvalues lie in \((-1, 1]\), and the eigenvalue 1 belongs to

\[ \hat A\, \mathbf v = \mathbf v, \qquad v_i \propto \sqrt{\tilde d_i} \tag{2}\]

which follows from \(\sum_j \hat A_{ij}\sqrt{\tilde d_j} = \sum_{j \in \{i\} \cup N(i)} 1/\sqrt{\tilde d_i} = \sqrt{\tilde d_i}\). Repeated layers pull every map towards \(\mathbf v\): they keep slow spatial variation and remove fast variation, a low-pass filter on the mesh (the mesh Laplacian of m06 in another form). A few layers average noise away; too many blur small targets.

\(\mathbf v\) is constant only when every vertex has the same degree. The course’s grid mesh alternates interior points with 4 and 8 neighbours, so its \(\mathbf v\) is a fine checkerboard with values in the ratio \(\sqrt{5} : \sqrt{9}\), and that pattern appears in the right panel of the demo below after a few layers. It comes from the normalization, not from noise. The row-normalized operator \(\tilde D^{-1}(A + I)\) maps a constant map to itself; the symmetric form is the one Kipf and Welling (2017) used.

Exact

Assumptions

  • Each layer applies \(\hat A\) of Equation 1 with no learned weights, exactly. The mesh is the course’s triangulated grid (epml.mapping.plane_mesh), 31 by 31 points 2 mm apart.
  • The map holds 1.0 outside a disc and 0.2 inside, plus noise with a fixed random pattern. These are illustrative values in arbitrary units, not voltages.
  • The fine checkerboard in the right panel is the fixed point of Equation 2 on this mesh, whose points have 4 or 8 neighbours.
  • What would change the result: learned weights choose what to keep; real meshes have uneven point density, so “L edges” means different distances in different places.

Key takeaways

  • A graph-convolution layer is a weighted neighbour average followed by a shared channel mix.
  • After \(L\) layers a vertex sees its \(L\)-edge neighbourhood.
  • Stacked layers low-pass filter the map: they remove noise and, eventually, small targets.

Judging a predicted region

A model that marks part of a map as a target can be judged by overlap or by location. The Dice coefficient of a predicted set \(P\) and a true set \(T\), with area weights on mesh vertices, is

\[ \mathrm{Dice} = \frac{2\,\lvert P \cap T \rvert}{\lvert P \rvert + \lvert T \rvert} \tag{3}\]

The centroid distance is the distance between the centres of \(P\) and \(T\). For two discs of radii \(R\) and \(r\) whose centres are \(d\) apart, the overlap has a closed form, the area of a lens:

\[ \lvert P \cap T\rvert = r^2 \cos^{-1}\!\frac{d^2 + r^2 - R^2}{2dr} + R^2 \cos^{-1}\!\frac{d^2 + R^2 - r^2}{2dR} - \tfrac12\sqrt{(-d + r + R)(d + r - R)(d - r + R)(d + r + R)} \tag{4}\]

for \(\lvert R - r\rvert < d < R + r\); the overlap is the smaller disc when \(d \le \lvert R - r\rvert\), and zero when \(d \ge R + r\). With two discs of radius 4 mm, centres 7 mm apart overlap by less than 15%. A small target can be “found” by centroid distance and missed by Dice. Report both.

Exact

Assumptions

  • Model: Equation 4 and Equation 3 for two discs on a plane, exact.
  • The default distance, 7.3 mm, echoes the median centroid distance reported by Saluja et al. (2024); the radii are illustrative and are not taken from that study.
  • What would change the result: real targets are not discs, and they lie on curved surfaces, where distances are geodesic (m06).

Key takeaways

  • Dice measures overlap; centroid distance measures location. They can disagree.
  • For small targets, a few millimetres of offset destroys overlap.
  • Choose the metric that matches the use: where to look, or what to treat as the target.

Rare targets

When targets cover a few percent of a map, a model trained with plain cross-entropy can do well on average by predicting “no target” almost everywhere. Two common fixes change the loss. A weight on positive examples scales their cross-entropy terms. The focal loss also shrinks the loss of examples the model already gets right (Lin et al. 2017):

\[ \mathrm{FL}(p_t) = -\alpha_t\,(1 - p_t)^{\gamma}\,\log p_t \tag{5}\]

where \(p_t\) is the predicted probability of the true class, \(\alpha_t = \alpha\) for positives and \(1 - \alpha\) for negatives, and \(\gamma \ge 0\). With \(\gamma = 0\) it is weighted cross-entropy; with \(\gamma > 0\), an example with \(p_t\) near 1 contributes almost nothing. Guler et al. (2026) used it with \(\alpha = 0.75\) and \(\gamma = 2\).

A loss changes what the model learns, not how it should be judged. With rare targets, report precision and recall (or the area under the precision–recall curve) rather than only ROC AUC: precision is the PPV of m35 at a low prevalence.

Key takeaways

  • Rare targets push a model towards “never”; weighting or a focal loss pushes back.
  • ROC AUC can look high while precision is low; quote precision and recall.
  • The operating threshold, not the loss, decides how many false targets a user sees.

Structure without labels

Some questions have no labels at all, for example whether a cohort’s voltage distributions fall into groups. A Gaussian mixture models values as coming from \(K\) normal components with weights \(\pi_k\), means \(\mu_k\) and standard deviations \(\sigma_k\). The probability that value \(x_i\) came from component \(k\) is its responsibility:

\[ \gamma_{ik} = \frac{\pi_k\, \mathcal N(x_i \mid \mu_k, \sigma_k)}{\sum_{j=1}^{K} \pi_j\, \mathcal N(x_i \mid \mu_j, \sigma_j)} \tag{6}\]

Expectation–maximization alternates between computing \(\gamma_{ik}\) and re-estimating each component from the values weighted by its responsibilities. Huttelmaier et al. (2026) used Gaussian mixtures in this way on left atrial voltage distributions. Clusters are a description of the data; whether they mean anything for patients needs outside evidence, as that study sought with other markers.

Key takeaways

  • A mixture assigns each value a probability of belonging to each group.
  • Expectation–maximization fits it without labels.
  • Clusters need external validation before they mean anything clinical.

Clinical correlate

NotePending clinical review

This section describes published findings only. It is not guidance for patient care, and it has not yet been reviewed by an EP physician.

AI-guided ablation of electrogram dispersion. Spatiotemporal dispersion was first described as a visually recognizable pattern of electrograms on a multispline catheter, which simulations and optical mapping linked to the vicinity of AF drivers (Seitz et al. 2017). That study compared dispersion- guided ablation with a separate validation set treated conventionally, not with a randomized control.

TAILORED-AF tested an AI algorithm that detects dispersion in a multicentre, randomized, double-blind trial in drug-refractory persistent AF (Deisenhofer, Albenque, et al. 2025). Against expert annotations, the algorithm had a ROC AUC of 0.94, reported in the trial’s supplement. Patients were randomized to ablation of dispersion areas plus PVI (n = 187) or PVI alone (n = 183). At 12 months after one procedure, freedom from documented AF was 88% versus 70%. Freedom from any atrial arrhythmia did not differ significantly between arms, procedure and ablation times were about twice as long in the tailored arm, and the composite safety endpoint did not differ. Later analyses looked at women (Deisenhofer, Seitz, Nguyen-Tu, et al. 2025) and at the atrial tachycardias seen after the procedure (Deisenhofer, Seitz, Lotteau, et al. 2025). A single-arm study applied the approach to repeat procedures in patients whose veins remained isolated (Hummel et al. 2026).

In the terms of this module and m45:

  • The randomized comparison tests the whole strategy (AI detection plus extra ablation plus PVI) against PVI, not the algorithm alone.
  • Endpoints differ in what they show. The primary endpoint (AF) favoured the tailored arm; the broader endpoint (any atrial arrhythmia) did not differ.
  • Single-arm results depend on selection. In the repeat-procedure study, many enrolled patients left before the primary endpoint, for example because their veins had reconnected (Hummel et al. 2026).

Visual summary

Schematic

Electrogramsat mapped points Featuresper point Graph layers X W Probability mapthreshold Target regionDice, centroid labels from expertsor another modality window vs patientsplits L layers seeL edges rare targets:weights, focal loss overlap andlocation differ A trial tests the whole strategy built on the map, not the network alone
Figure 1: From electrograms to a predicted target region, with the question to ask at each stage (gray).

Assumptions

How experts think

Before trusting a model that reads electrograms or maps, ask:

  1. Where did the labels come from? Expert annotation (with its agreement, (Alhusseini et al. 2020)), another modality such as optical mapping (Zolotarev et al. 2020), or the outcome of an ablation?
  2. What was split? Windows from one recording, points from one map and maps from one patient are all correlated. Split by patient (lab part E).
  3. How far does the model look? A graph network with \(L\) layers sees \(L\) edges; a per-point model sees one point (Equation 1).
  4. Which metric matches the use? Dice for “this is the target”, centroid distance for “look here” (Equation 3, Equation 4), precision and recall for rare targets (Equation 5).
  5. What did the trial test? In TAILORED-AF, the comparison was a strategy built on the AI, against PVI alone (Deisenhofer, Albenque, et al. 2025).

One sentence to keep: on a map, every point has neighbours; models, splits and metrics must know it.

Cheat sheet

Quantity Result
Graph convolution (Equation 1) \(H' = \sigma(\hat A H W)\), \(\hat A = \tilde D^{-1/2}(A+I)\tilde D^{-1/2}\)
Smoothing fixed point (Equation 2) \(\hat A\mathbf v = \mathbf v\) for \(v_i \propto \sqrt{\tilde d_i}\); eigenvalues in \((-1, 1]\)
Dice (Equation 3) \(2\lvert P\cap T\rvert/(\lvert P\rvert + \lvert T\rvert)\)
Lens area (Equation 4) Overlap of two discs, radii \(R\), \(r\), centres \(d\) apart
Focal loss (Equation 5) \(-\alpha_t(1-p_t)^\gamma\log p_t\); \(\gamma = 0\) gives weighted cross-entropy
Mixture responsibility (Equation 6) \(\gamma_{ik} = \pi_k\mathcal N_k(x_i)/\sum_j \pi_j\mathcal N_j(x_i)\)

Studies in this module

Study What was learned Evidence
Alhusseini et al. (2020) Rotational activation in AF activation grids Retrospective; separate test set
Zolotarev et al. (2020) Driver electrodes, labeled by optical mapping Ex vivo human atria
Saluja et al. (2024) Isthmus areas of scar-related atrial tachycardia (graph network) Retrospective, 29 tachycardias
Guler et al. (2026) Ganglionated plexi from bipolar electrograms Retrospective, held-out atria
Deisenhofer, Albenque, et al. (2025) Strategy: AI-detected dispersion ablation plus PVI vs PVI Randomized trial

Rules of thumb

  • Split by patient; windows and points from one patient are not independent.
  • Two to three graph layers smooth noise; more layers blur small targets.
  • For rare targets, show the precision–recall curve.

Interview questions

Try each question before you open its model answer. In a real interview, say which assumptions you are making.

Q1. Write one graph-convolution layer and explain each factor.

Model answer

\(H' = \sigma(\hat A H W)\) with \(\hat A = \tilde D^{-1/2}(A + I)\tilde D^{-1/2}\) (Equation 1). \(A + I\) links each vertex to its neighbours and itself; the degree normalization keeps high-degree vertices from dominating; \(W\) mixes feature channels and is shared by all vertices, so the layer works on meshes of any size; \(\sigma\) adds nonlinearity. After \(L\) layers a vertex sees its \(L\)-edge neighbourhood.

Q2. Why does stacking many graph layers wash out a map?

Model answer

\(\hat A\) is symmetric with eigenvalues in \((-1, 1]\), and its eigenvector for 1 is \(v_i \propto \sqrt{\tilde d_i}\) (Equation 2). Repeated application shrinks every other component, so the map tends towards that one shape: high spatial frequencies go first. Noise is removed, but so are small targets. Demo m39-d1 shows both.

Q3. A model’s predicted target centre is 7 mm from the true centre. Is that good?

Model answer

It depends on the target’s size and the use. For two discs of radius 4 mm, 7 mm apart, Dice is below 15% (Equation 4). Saluja et al. (2024) reported a median centroid distance of 7.3 mm with a mean Dice of 13.9%. As a pointer to where to map more densely it may help; as a definition of what to ablate it overlaps the target poorly.

Q4. Your positive class is 1% of points. The model’s ROC AUC is 0.94. What else do you report?

Model answer

Precision and recall at the chosen threshold, and the precision–recall curve or its area. At low prevalence, a modest false-positive rate swamps the true positives, so precision can be low while ROC AUC looks high; precision is PPV at that prevalence (m35). Guler et al. (2026) is an example: recall 0.85, precision 0.09.

Q5. What does the focal loss change, and what does it not change?

Model answer

It down-weights examples the model already classifies well, by \((1-p_t)^\gamma\), and can weight classes by \(\alpha\) (Equation 5). Training then concentrates on hard and rare examples. It does not change how the model should be evaluated, and it does not choose the operating threshold; that is a separate decision about acceptable false positives.

Q6. You have 30-second recordings from 8 patients, 4 catheter positions each, cut into 4-s windows. How do you split for testing?

Model answer

By patient: leave one patient out, train on the others, and rotate. Windows from one recording share the patient, the catheter contact and the rhythm, so a window-wise split lets the model recognize the recording instead of the pattern. Lab part E shows the gap on the Intracardiac AF Database.

Q7. Why did averaging features over neighbouring electrodes help driver detection?

Model answer

A driver is a spatial pattern: neighbouring electrodes share evidence for it, while noise at single electrodes is partly independent. Averaging over the 8 neighbours (Zolotarev et al. 2020) raises the signal relative to the noise, the same idea as one unweighted graph-convolution layer (Equation 1).

Q8. In TAILORED-AF, what exactly was compared?

Model answer

Two strategies in drug-refractory persistent AF: ablation of AI-detected dispersion areas plus PVI, against PVI alone, randomized and double-blind (Deisenhofer, Albenque, et al. 2025). The primary endpoint, freedom from documented AF at 12 months after one procedure, was 88% versus 70%; freedom from any atrial arrhythmia did not differ significantly, and procedures took about twice as long in the tailored arm. The trial tests the strategy, not the algorithm’s accuracy on its own.

Q9. What does a Gaussian mixture give you that a fixed voltage threshold does not?

Model answer

A description of the whole distribution: how many groups, their means and spreads, and for each value a probability of belonging to each group (Equation 6). It does not depend on a threshold chosen in advance (Huttelmaier et al. 2026). It does not tell you that the groups matter clinically; that needs outside evidence.

Q10. A paper reports 95% accuracy for classifying activation grids. What do you check?

Model answer

How labels were made and how much experts agreed (κ = 0.79 in Alhusseini et al. (2020)); whether test grids came from patients not used in training; class balance; what “95%” counts (grids, not patients); whether the explanation of the model matched expert rules; and whether anything beyond accuracy was tested, such as use during procedures.

Q11. Why is a graph network a natural fit for mapping data, compared with an image network?

Model answer

Mapping points sit on a curved surface with uneven density; there is no regular pixel grid. A graph network works directly on the mesh’s connections, shares weights across vertices and handles any number of points. An image network needs the surface flattened or resampled, which distorts distances. The lab compares a graph network with a per-point model on the same features.

Q12. Name two ways the AI-guided ablation studies could mislead if read carelessly.

Model answer

Reading a strategy trial as proof that the algorithm finds the true mechanism, when it tests the whole approach against PVI (Deisenhofer, Albenque, et al. 2025). And reading a single-arm study’s success rate without its exclusions: in the repeat-procedure study many enrolled patients left before the endpoint, for example with reconnected veins (Hummel et al. 2026), so the rate applies to those who remained.

Code lab

The lab builds graph layers on a mesh, trains a graph network to find slow regions on synthetic maps, deals with rare targets, and then checks a placement classifier on real intracardiac recordings with honest splits. It runs in under a minute, plus downloads.

Part What you do Data
A Build \(\hat A\) on a mesh; check Equation 1 and Equation 2; smooth a noisy map synthetic
B Synthetic mapping cases with slow regions; per-point features synthetic
C A per-point model against a NumPy graph network; ROC AUC, PR AUC, Dice and centroid distance synthetic
D Rare targets: cross-entropy, weighted cross-entropy and focal loss (Equation 5) synthetic
E Catheter position from electrogram features: window-wise against leave-one-patient-out PhysioNet
F Optional: the graph network in PyTorch synthetic

Each part ends with checks: lines that pass with the code as given. If you change something and a check fails, that is information, not an error to hide.

Self-check

Short checks you can do in your head. Open each answer after you have tried it.

S1. A vertex has 6 neighbours, and each neighbour also has 6. What is \(\hat A_{ij}\) for a neighbour, and for the vertex itself?

Answer

\(\tilde d = 7\) for all of them, so \(\hat A_{ij} = 1/7\) for each neighbour and for itself (Equation 1): a plain average of 7 values.

S2. After 3 layers on a grid with 2-mm spacing, how far along the grid lines can information travel?

Answer

3 edges, so 6 mm along grid lines (diagonal edges reach a little further).

S3. Prediction and truth are the same disc. Dice?

Answer

1 (Equation 3).

S4. Two discs of radius 3 mm, centres 6 mm apart. Dice?

Answer

0: they touch at one point, since \(d = R + r\) (Equation 4).

S5. With \(\gamma = 0\) and \(\alpha = 0.5\), what is the focal loss?

Answer

Half the ordinary cross-entropy (Equation 5).

S6. True or false: a window-wise split and a patient-wise split estimate the same thing.

Answer

False. The window-wise split estimates performance on new windows from patients already seen; the patient-wise split estimates performance on new patients, which is what matters for use.

Literature figures

These figures are worth looking at alongside this module. We link to them instead of copying them: some licenses do not allow reuse here, and others we could not confirm. Each entry says what to look for and where this module explains it.

  • What the network looked at. Alhusseini and colleagues, the gradient-weighted class activation (Grad-CAM) heatmaps, which they compared with the consensus of expert raters [Alhusseini et al. (2020); free full text]. This module: From the field; interview Q10.
  • A neighbourhood of electrodes. Zolotarev and colleagues, the figure showing how features are averaged over the 8 electrodes around each electrode of a 3-by-3 grid [Zolotarev et al. (2020); free full text]. Compare with one unweighted graph layer (Equation 1, interview Q7).
  • Two endpoints, two answers. Deisenhofer and colleagues, the Kaplan–Meier curves for freedom from AF and from any atrial arrhythmia [Deisenhofer, Albenque, et al. (2025); free full text]. Compare the two endpoints (clinical correlate).

Further reading

  • Methods. Kipf and Welling (2017) (graph convolutional networks), Lin et al. (2017) (focal loss).
  • Electrogram learning. Alhusseini et al. (2020), Zolotarev et al. (2020), Guler et al. (2026), Rodrigo et al. (2022) (m38).
  • Maps. Saluja et al. (2024) (graph network), Huttelmaier et al. (2026) (mixtures on voltage).
  • Outcome prediction. Tang et al. (2022) and Ganesan et al. (2025), from Stanford registries.
  • Dispersion and trials. Seitz et al. (2017), Deisenhofer, Albenque, et al. (2025) and its follow-ups (Deisenhofer, Seitz, Nguyen-Tu, et al. 2025; Deisenhofer, Seitz, Lotteau, et al. 2025), Hummel et al. (2026).
  • AI competency areas. Machine learning; graphical models (graph structure on the mapping mesh, without a probabilistic model). See the AI competency map.

References

Alhusseini, Mahmood I., Firas Abuzaid, Albert J. Rogers, et al. 2020. “Machine Learning to Classify Intracardiac Electrical Patterns During Atrial Fibrillation: Machine Learning of Atrial Fibrillation.” Circulation: Arrhythmia and Electrophysiology 13: e008160. https://doi.org/10.1161/CIRCEP.119.008160.
Deisenhofer, Isabel, Jean-Paul Albenque, Sonia Busch, et al. 2025. “Artificial Intelligence for Individualized Treatment of Persistent Atrial Fibrillation: A Randomized Controlled Trial.” Nature Medicine 31 (4): 1286–93. https://doi.org/10.1038/s41591-025-03517-w.
Deisenhofer, Isabel, Julien Seitz, Sabine Lotteau, et al. 2025. “Targetable Atrial Tachycardias After Artificial Intelligence-Guided Ablation of Persistent Atrial Fibrillation.” Heart Rhythm 23 (3): e357–67. https://doi.org/10.1016/j.hrthm.2025.11.018.
Deisenhofer, Isabel, Julien Seitz, Marie-Sophie Nguyen-Tu, et al. 2025. “Women with Persistent Atrial Fibrillation Need More Than Pulmonary Vein Isolation: Personalised Extra-Pulmonary Vein Ablation Strategy Vs. Pulmonary Vein Isolation Alone in the TAILORED-AF Trial.” Europace 27 (11). https://doi.org/10.1093/europace/euaf281.
Ganesan, Prasanth, Maxime Pedron, Ruibin Feng, et al. 2025. “Comparing Phenotypes for Acute and Long-Term Response to Atrial Fibrillation Ablation Using Machine Learning.” Circulation. Arrhythmia and Electrophysiology 18 (2): e012860. https://doi.org/10.1161/CIRCEP.124.012860.
Goldberger, A. L., L. A. Amaral, L. Glass, et al. 2000. “PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Research Resource for Complex Physiologic Signals.” Circulation 101 (23): e215–20. https://doi.org/10.1161/01.cir.101.23.e215.
Guler, Tumer Erdem, Metin Cagdas, Sukriye Ebru Onder, et al. 2026. “Convolutional Neural Network for Real‑time Localization of Ganglionated Plexi from Bipolar Intracardiac Electrograms.” Journal of Interventional Cardiac Electrophysiology 69 (5): 957–68. https://doi.org/10.1007/s10840-026-02307-9.
Hummel, John D, Paul C Zei, Mark Metzl, et al. 2026. “Reablation of Atrial Fibrillation Targeting Electrogram Dispersion in Patients with Isolated Veins: The RESTART Trial.” Heart Rhythm 23 (5): 1083–90. https://doi.org/10.1016/j.hrthm.2026.01.042.
Huttelmaier, Moritz T, Manuel Vogel, Jonas Herting, et al. 2026. “A Novel Signal Distribution-Based Approach to Characterize Atrial Cardiomyopathy.” Journal of Interventional Cardiac Electrophysiology 69 (5): 977–87. https://doi.org/10.1007/s10840-026-02313-x.
Intracardiac Atrial Fibrillation Database (Version 1.0.0). 2003. PhysioNet. https://doi.org/10.13026/C23S33.
Kipf, Thomas N., and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. arXiv:1609.02907. https://doi.org/10.48550/arXiv.1609.02907.
Lin, Tsung-Yi et al. 2017. Focal Loss for Dense Object Detection. arXiv:1708.02002. https://doi.org/10.48550/arXiv.1708.02002.
Rodrigo, Miguel, Mahmood I Alhusseini, Albert J Rogers, et al. 2022. “Atrial Fibrillation Signatures on Intracardiac Electrograms Identified by Deep Learning.” Computers in Biology and Medicine 145: 105451. https://doi.org/10.1016/j.compbiomed.2022.105451.
Saluja, Deepak, Ziyi Huang, Jonah Majumder, et al. 2024. “Automated Prediction of Isthmus Areas in Scar-Related Atrial Tachycardias Using Artificial Intelligence.” Journal of Cardiovascular Electrophysiology 35 (7): 1401–11. https://doi.org/10.1111/jce.16299.
Seitz, Julien, Clément Bars, Guillaume Théodore, et al. 2017. “AF Ablation Guided by Spatiotemporal Electrogram Dispersion Without Pulmonary Vein Isolation: A Wholly Patient-Tailored Approach.” Journal of the American College of Cardiology 69 (3): 303–21. https://doi.org/10.1016/j.jacc.2016.10.065.
Tang, Siyi, Orod Razeghi, Ridhima Kapoor, et al. 2022. “Machine Learning-Enabled Multimodal Fusion of Intra-Atrial and Body Surface Signals in Prediction of Atrial Fibrillation Ablation Outcomes.” Circulation. Arrhythmia and Electrophysiology 15 (8): e010850. https://doi.org/10.1161/CIRCEP.122.010850.
Zolotarev, Alexander M., Brian J. Hansen, Ekaterina A. Ivanova, et al. 2020. “Optical Mapping-Validated Machine Learning Improves Atrial Fibrillation Driver Detection by Multi-Electrode Mapping.” Circulation: Arrhythmia and Electrophysiology 13: e008249. https://doi.org/10.1161/CIRCEP.119.008249.