Skip to main content

m41 · Locating arrhythmia origins from the ECG: regression and deep learning, simulation-trained models, active pace-mapping

Warning

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

In 60 seconds

The QRS complex of a paced beat or of a premature ventricular beat carries information about where activation began. Regression and machine-learning models turn that information into a site on the heart. A population model fitted on 1,012 pacing sites from 38 patients placed test sites with a mean error of about 11 to 12.5 mm on a generic left ventricle (Zhou, AbdelWahab, et al. 2018); patient-specific models that learn from a few pacing sites during the procedure reached about 6 mm in prospective case series (Zhou et al. 2021); and active methods choose where to pace next (Missel et al. 2020; Meisenzahl et al. 2024). This module derives why a QRS integral is a weighted sum of activation times, fits localization by least squares, measures error along the heart surface, and builds the Gaussian-process and expected-improvement loop behind active pace-mapping.

From the field

  • Population, then patient. In one group’s work, a population regression on 120-ms QRS integrals gave mean test errors of 11.4 mm (support-vector regression) and 12.5 mm (linear regression) (Zhou, AbdelWahab, et al. 2018). With patient-specific coefficients learned from pacing during the procedure, the same group reported 7.2 ± 4.1 mm for 25 tachycardia exit sites after 3 remote and 5 close pacing sites (Zhou et al. 2020), and 5.9 ± 2.6 mm for 26 sites in a prospective series of patients with structural heart disease using three leads (III, V2 and V6) (Zhou et al. 2021).
  • A shell from CT instead of a full map. Registering a CT-derived ventricular surface allowed the same approach without a complete electroanatomic map: 6.9 ± 5.7 mm for 20 tachycardias (Zhou et al. 2019). In one patient, it was compared with ECG imaging from 120 body-surface leads: 8.9 against 24.9 mm on 14 pacing sites (Zhou, Horáček, et al. 2018) (ECG imaging: m32).
  • Choosing the next pacing site. A hybrid model trained on 38 patients and tested on a separate cohort of 8 localized left-ventricular pacing sites to 5.3 ± 2.6 mm using 5.4 ± 2.5 pacing sites (Missel et al. 2020). A Bayesian-optimization method that models ECG similarity over the heart surface, and can avoid suggesting sites inside scar, reached 3.9 ± 3.6 mm with 8.0 ± 4.0 sites in simulation (Meisenzahl et al. 2024).
  • Training on simulations. A classifier of outflow-tract origin (right or left) trained on 2,496 simulated ECGs from 16 patient geometries reached accuracies of 0.86 and 0.84 on two clinical data sets, against 0.84 and 0.76 when trained on real ECGs alone; lead V2 carried the most information (Doste et al. 2022). One of those test sets is an open database of 334 patients whose origin was set by the site of successful ablation (Zheng et al. 2020).
  • Small cohorts need intervals. A logistic regression on ECG features, age and sex in 43 patients separated right from left outflow-tract origin with a pooled ROC AUC of 0.910 (95% CI 0.800 to 0.987) (Botis et al. 2026). Reviews summarize the field and its open problems, such as external validation (Yang et al. 2024; Antoun et al. 2026).

Derivations

Symbols: electrodes \(e\), surface elements \(j\) with activation times \(\tau_j\), solid angle \(\Omega_{ej}\) of element \(j\) seen from electrode \(e\), potential scale \(c\), integration window \(T\), feature vector \(\mathbf x\) (one QRS integral per lead), origin coordinates \(\mathbf r\), kernel length scale \(\ell\), signal scale \(\sigma_f\), noise \(\sigma_n\), posterior mean \(\mu\) and standard deviation \(\sigma\).

Why a QRS integral points to the origin

A simple model of ventricular activation puts a uniform double layer on the part of a closed ventricular surface that has activated. In an infinite homogeneous medium, an electrode then sees a potential proportional to the solid angle of the activated part (volume conduction: m02): \(\phi_e(t) = c \sum_j \Omega_{ej}\, H(t - \tau_j)\), with \(H\) the unit step. Integrating over the first \(T\) milliseconds,

\[ x_e = \int_0^T \phi_e(t)\, dt = c \sum_j \Omega_{ej}\,(T - \tau_j)_+ = -c \sum_j \Omega_{ej}\, \tau_j \quad \text{when } T \ge \max_j \tau_j \tag{1}\]

The last step uses \(\sum_j \Omega_{ej} = 0\): from outside a closed surface, its solid angles cancel. So a lead’s QRS integral is a fixed, lead-specific weighted sum of activation times. The activation times depend on the origin through distances along the heart, so the map from origin to features is smooth but not linear. Zhou and colleagues used the 120-ms integral of each lead as the feature (Zhou, AbdelWahab, et al. 2018); lab part A checks Equation 1 against numerical integration.

Key takeaways

  • A QRS integral weights activation times by how each lead sees each part of the heart.
  • The map from origin to integrals is smooth, so nearby origins give similar features.
  • The model ignores the torso boundary, wall thickness and repolarization; it is a tool for intuition.

Fitting a localization model

Collect features \(\mathbf x_i\) and known origins \(\mathbf r_i\) (pacing sites) and fit \(\mathbf r \approx W^\top \mathbf x + \mathbf b\) by least squares. With \(A = [X \;\; \mathbf 1]\),

\[ \begin{bmatrix} W \\ \mathbf b^\top \end{bmatrix} = (A^\top A)^{-1} A^\top R \tag{2}\]

A population model is fitted once on many patients and maps features to a generic heart; a patient-specific model is fitted during the procedure on that patient’s own pacing sites. With \(p\) features the fit needs at least \(p + 1\) sites: 4 for three leads. Errors fall fastest when later sites are close to the current estimate, because the map is close to linear over small regions (lab part C) (Zhou et al. 2020).

Measuring error on the heart

Localization error is a distance along the heart surface, not through the cavity. On a sphere of radius \(R\), for unit vectors \(\mathbf u\) and \(\mathbf v\),

\[ d = R \arccos(\mathbf u \cdot \mathbf v), \qquad \lvert \mathbf u - \mathbf v \rvert R = 2R\sin\frac{d}{2R} \le d \tag{3}\]

so a straight-line error understates the distance along the wall, by more the farther apart the points are. On a mesh, geodesic distance is computed as in m06; zhou2018 measured it on a generic left-ventricular surface. The lab uses shortest paths along mesh edges, which are on average about 6% longer than true geodesics on its mesh.

Key takeaways

  • Least squares needs at least one more site than features.
  • Population models generalize across patients; patient-specific models adapt to one.
  • Report error along the surface, and say how it was computed.

Choosing the next pacing site

Pace mapping compares the ECG of a paced beat with the target beat. A common score is the mean correlation over the 12 leads:

\[ \rho = \frac{1}{12} \sum_{l=1}^{12} \frac{\sum_t (a_l(t) - \bar a_l)(b_l(t) - \bar b_l)}{\sqrt{\sum_t (a_l(t) - \bar a_l)^2 \sum_t (b_l(t) - \bar b_l)^2}} \tag{4}\]

It ignores amplitude and offset per lead. The score \(f(\mathbf s)\) as a function of the pacing site \(\mathbf s\) is unknown but smooth. A Gaussian process with a squared-exponential kernel \(k(\mathbf s, \mathbf s') = \sigma_f^2 \exp(-\lvert\mathbf s - \mathbf s'\rvert^2/2\ell^2)\), conditioned on the scores at the paced sites \(S\), gives a posterior at any site (Frazier 2018):

\[ \mu(\mathbf s) = m + \mathbf k_*^\top (K + \sigma_n^2 I)^{-1} (\mathbf f - m), \qquad \sigma^2(\mathbf s) = \sigma_f^2 - \mathbf k_*^\top (K + \sigma_n^2 I)^{-1} \mathbf k_* \tag{5}\]

where \(K\) holds the kernel between paced sites, \(\mathbf k_*\) the kernel between \(\mathbf s\) and the paced sites, \(\mathbf f\) the observed scores and \(m\) a constant prior mean. Expected improvement over the best score so far, \(f^*\), has a closed form (Frazier 2018):

\[ \mathrm{EI}(\mathbf s) = \Delta\, \Phi\!\left(\frac{\Delta}{\sigma}\right) + \sigma\, \varphi\!\left(\frac{\Delta}{\sigma}\right), \qquad \Delta = \mu(\mathbf s) - f^* - \xi \tag{6}\]

with \(\Phi\) and \(\varphi\) the standard normal distribution and density and \(\xi \ge 0\) a margin; when \(\sigma = 0\), \(\mathrm{EI} = \max(\Delta, 0)\). It is large where the mean is high (exploitation) or the uncertainty is high (exploration). Pacing where EI peaks, then updating, is the loop behind methods such as Meisenzahl et al. (2024), which add constraints such as avoiding scar.

Simulation

Assumptions

  • Model: a Gaussian process with a squared-exponential kernel (Equation 5; signal scale 0.5, noise 0.01, length scale on the slider) and expected improvement (Equation 6) on a 31-by-31 grid of candidate sites 2 mm apart. The pacing sequence is recomputed exactly for each setting.
  • The similarity landscape is illustrative: a main peak at the origin and a weaker peak elsewhere, standing in for a site whose paced QRS looks similar. It is not computed from an ECG model.
  • Every sequence starts from the same two corner sites; the random rule uses a fixed seed.
  • What would change the result: real similarity is noisy and depends on the leads and the window; sites inside scar cannot be paced; the surface is curved, not flat.

Exact

Assumptions

  • Exact: Equation 5 with a squared-exponential kernel and noise 0.01, prior mean equal to the mean of the four scores, and Equation 6, evaluated on 241 points along the line.
  • The four positions and scores are illustrative.
  • What would change the result: hyperparameters are usually fitted to the data, not set by hand; with noisy scores the best observed value is itself uncertain (Frazier 2018).

Key takeaways

  • Pace-map correlation compares shapes, not amplitudes.
  • A Gaussian process gives both an estimate and its uncertainty at every candidate site.
  • Expected improvement trades a high estimate against high uncertainty in one closed formula.

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.

Where these tools sit. Before an ablation for premature ventricular beats or ventricular tachycardia, the 12-lead ECG is used to estimate the region of origin; during the procedure, activation mapping and pace mapping refine it (m11, m24) (Zheng et al. 2020; Antoun et al. 2026). The methods in this module aim to make the estimate quantitative: a site on the patient’s surface with an error in millimetres.

What the evidence is. The results cited here come from retrospective and prospective case series (Zhou et al. 2020, 2019, 2021), a separate test cohort (Missel et al. 2020), simulation (Meisenzahl et al. 2024), and retrospective classification with cross-validation (Doste et al. 2022; Botis et al. 2026). Among them we found no randomized comparison of procedures with and without such a tool. A review reports early data suggesting shorter procedures and less fluoroscopy (Antoun et al. 2026).

What “ground truth” means. In the open outflow-tract database, the origin is the site of successful ablation, and only patients with a single source were included (Zheng et al. 2020). In pacing studies, the reference is the catheter position recorded by the mapping system. Both are good references with limits: a lesion can succeed at a site near, not at, the origin, and excluding multiple sources makes the task easier than the clinic.

In the terms of this module and m45:

  • Read the error metric. A mean error in millimetres along the surface, with its spread, says more than “accurate”.
  • Read the reference. Pacing sites, ablation success and activation mapping give different ground truths.
  • Read the cohort. Results on simulated hearts, on one centre’s patients or on a separate cohort answer different questions.

Visual summary

Schematic

Target ECGPVC or VT beat FeaturesQRS integrals Population modelfirst estimate Site on the patient'ssurface, error in mm Pace, compare, updateregression or GP + EI reference: pacingor ablation success? which leads,which window? along the surface,with its spread The clinician decides where to pace; the model only suggests
Figure 1: A population model gives a first estimate; a patient-specific or active loop refines it with each pacing site, with the question to ask at each stage (gray).

Assumptions

How experts think

Before trusting a model that places an origin from the ECG, ask:

  1. What was the reference? Pacing coordinates, the site of successful ablation, or activation mapping (Zheng et al. 2020; Zhou et al. 2021).
  2. What was split? Pacing sites from one patient are correlated; a split by patient tests new patients (lab part B).
  3. How was error measured? Along the surface or straight through, on a generic or the patient’s own geometry (Equation 3).
  4. How many pacing sites did it need, and where were they? Patient-specific and active methods trade pacing effort for accuracy (Equation 2; Equation 6).
  5. Simulation or patients? Simulation-trained models need testing on real ECGs from other centres (Doste et al. 2022); simulation results (Meisenzahl et al. 2024) are a target, not a clinical result.

One sentence to keep: an origin estimate is only as good as its reference, its split and its distance metric.

Cheat sheet

Quantity Result
QRS integral (Equation 1) \(x_e = c\sum_j \Omega_{ej}(T-\tau_j)_+\); \(= -c\sum_j\Omega_{ej}\tau_j\) after full activation
Least squares (Equation 2) \([W; \mathbf b^\top] = (A^\top A)^{-1}A^\top R\); needs \(\ge p + 1\) sites
Surface error (Equation 3) \(d = R\arccos(\mathbf u\cdot\mathbf v) \ge\) chord \(2R\sin(d/2R)\)
Pace-map similarity (Equation 4) Mean Pearson correlation over 12 leads
GP posterior (Equation 5) \(\mu = m + \mathbf k_*^\top(K+\sigma_n^2I)^{-1}(\mathbf f - m)\); \(\sigma^2 = \sigma_f^2 - \mathbf k_*^\top(K+\sigma_n^2I)^{-1}\mathbf k_*\)
Expected improvement (Equation 6) \(\Delta\Phi(\Delta/\sigma) + \sigma\varphi(\Delta/\sigma)\), \(\Delta = \mu - f^* - \xi\)

Studies in this module

Study What was done Evidence
Zhou, AbdelWahab, et al. (2018) Population regression, generic left ventricle 1,012 pacing sites, 38 patients, random 80/20 split
Zhou et al. (2021) Patient-specific regression, three leads Retrospective and prospective case series, 42 patients
Missel et al. (2020) Hybrid population and active model Separate test cohort of 8 patients
Meisenzahl et al. (2024) Bayesian optimization of pacing sites Simulation
Doste et al. (2022) Classifier trained on simulated ECGs Two clinical test sets, 365 patients
Botis et al. (2026) Right vs left outflow tract, logistic regression Retrospective, 43 patients, cross-validation

Rules of thumb

  • Split by patient; pacing sites from one heart are not independent.
  • Pair a population model with patient-specific updates when pacing is possible.
  • Quote errors along the surface, with spread, and the number of pacing sites used.

Interview questions

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

Q1. Why does the integral of the QRS carry information about the origin?

Model answer

In a double-layer model, each lead’s potential is the solid angle of the activated surface, so the integral over the QRS is a lead-specific weighted sum of activation times (Equation 1). Activation times depend on distance from the origin, so each origin produces its own pattern of integrals across leads. The relation is smooth, which is why regression works, and nonlinear, which is why patient-specific and local fits help.

Q2. Population model or patient-specific model?

Model answer

A population model works before any pacing and needs no patient data, but it carries the spread of anatomy and lead placement between patients: mean errors around 11 to 12.5 mm in Zhou, AbdelWahab, et al. (2018). A patient-specific model learns that patient’s geometry and leads from a few pacing sites and has reached about 6 mm (Zhou et al. 2021), at the cost of pacing time. In practice the first gives a starting region and the second refines it (lab part C).

Q3. With three leads, what is the smallest number of pacing sites for a patient-specific linear model, and why?

Model answer

Four: three coefficients per coordinate plus an intercept (Equation 2). With exactly four the fit passes through every site and has no redundancy against noise, so more are used in practice; the cited system used at least five (Zhou et al. 2019).

Q4. Why report error along the surface rather than straight-line distance?

Model answer

Catheters and ablation act on the wall; two points on opposite walls can be close in straight lines but far along the surface. The chord always understates the surface distance (Equation 3), increasingly for distant points. State how the surface distance was computed (mesh geodesic, edge paths) and on which geometry.

Q5. A method localizes origins to 3.9 mm in simulation. How do you read that?

Model answer

As a target under ideal conditions (Meisenzahl et al. 2024): the simulated ECGs come from the same model family as the training, noise and lead placement are controlled, and the reference is known exactly. Check the spread (± 3.6 mm there), the number of pacing sites (8.0 ± 4.0), the geometries and tissue properties covered, and whether a clinical cohort was tested.

Q6. What does the Gaussian-process length scale control in active pace-mapping?

Model answer

How far a measurement informs the estimate (Equation 5). A short length scale treats sites as nearly independent, so the search explores widely; a long one smooths over the surface and can jump to a wrong peak with confidence. Demo m41-d1 shows both. Hyperparameters are usually fitted, with care, since few points are available.

Q7. Explain expected improvement in one sentence, and say what happens where the model is certain.

Model answer

It is the average amount by which pacing at a site would beat the best score so far, given the model’s estimate and uncertainty (Equation 6). Where \(\sigma = 0\) it reduces to \(\max(\mu - f^*, 0)\), which is zero at sites already paced: the rule never repeats a site without a reason.

Q8. What can make a pace-map correlation misleading?

Model answer

Correlation ignores amplitude and offset (Equation 4), so two different sites can give similar shapes; the score depends on the leads, their positions and the time window; and noise or a changed heart rate between the target and the paced beat alters it. Demo m41-d1 includes a second, weaker peak to show how a search can be drawn to a similar-looking site.

Q9. In lab part B, the random split by pacing site gives a lower error than the split by patient. Why?

Model answer

With a random split, sites from every patient are in training, so the model has seen each test patient’s geometry and lead positions. That estimates performance on new sites in known patients, not on new patients. Note the design of each study: Zhou, AbdelWahab, et al. (2018) split its 1,012 sites at random.

Q10. A model trained on simulated ECGs is applied to patients. What can go wrong, and how was it tested?

Model answer

Simulations cover a limited range of anatomy, lead positions and tissue properties, so real ECGs may fall outside it (lab part E). Doste et al. (2022) tested on two clinical data sets from other hospitals and reported accuracies of 0.86 and 0.84, better than training on real ECGs alone; that is the right kind of test, and it still covers only outflow-tract origins.

Q11. A 43-patient model has a ROC AUC of 0.910 with a 95% CI of 0.800 to 0.987. What do you conclude?

Model answer

That discrimination is probably good but imprecisely known: the interval is wide (Botis et al. 2026). Adding age and sex to ECG features changed the AUC by 0.057, not significantly (p = 0.104). It needs external validation before it can be compared with established ECG criteria.

Q12. How does learned localization differ from ECG imaging?

Model answer

ECG imaging (m32) solves an inverse problem with a torso and heart geometry and many body-surface leads, and returns a whole map. Learned localization maps a few features from standard leads to one site, using data from many patients or from pacing in this patient. In one patient, the pacing-based regression was more accurate than ECG imaging for pacing sites (Zhou, Horáček, et al. 2018); one case cannot rank the methods in general.

Code lab

The lab builds a double-layer forward model of a paced ventricle, fits population and patient-specific localization models, runs an active pace-mapping loop with a Gaussian process, and shows the cost of training hearts that differ from test hearts. It runs in about a minute.

Part What you do Data
A Forward model and twelve leads; check Equation 1 synthetic
B Population regression from twelve QRS integrals; split by patient or by site synthetic
C Patient-specific regression from three leads; random against sequential pacing synthetic
D Active pace-mapping: Gaussian process and expected improvement against random synthetic
E Training hearts that differ from test hearts synthetic
F Optional: a small neural network in PyTorch synthetic
  • Python notebook: labs/python/m41_lab.ipynb. Run it from the course repository after pip install -r requirements.txt, or open it in Google Colab.
  • Data: synthetic only. We found no open data set with pacing-site coordinates that we could verify for automatic download. The open outflow-tract database (Zheng et al. 2020) labels origins as right or left outflow tract and sub-regions; it is a candidate for a later classification part.
  • Part F runs only if PyTorch is installed.

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 patient-specific linear model uses four leads. How many pacing sites does it need at least?

Answer

Five: four coefficients plus an intercept (Equation 2).

S2. On a sphere of radius 30 mm, two sites are 90° apart. Distance along the surface, and in a straight line?

Answer

\(30 \times \pi/2 = 47.1\) mm along the surface; \(2 \times 30 \times \sin 45° = 42.4\) mm in a straight line (Equation 3).

S3. At a candidate site the posterior mean equals the best score so far and \(\sigma = 0.2\). What is the expected improvement (\(\xi = 0\))?

Answer

\(0.2\,\varphi(0) = 0.2/\sqrt{2\pi} \approx 0.080\) (Equation 6).

S4. A paced ECG equals the target ECG with every lead doubled. What is the pace-map correlation?

Answer

1: correlation ignores per-lead scale (Equation 4).

S5. In Equation 1, why does the window length \(T\) drop out once the whole surface has activated?

Answer

Its coefficient is \(c\sum_j\Omega_{ej}\), and the solid angles of a closed surface seen from outside sum to zero.

S6. True or false: a Gaussian process predicts the similarity score at an unpaced site with no uncertainty.

Answer

False. Away from paced sites the posterior variance grows towards \(\sigma_f^2\) (Equation 5); that uncertainty is what expected improvement uses.

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.

  • Errors per patient. Zhou and colleagues, Figure 1, the localization error for the pacing sites of each patient in the prospective series and the mean per patient [Zhou et al. (2021); free full text]. Compare the spread between patients with the mean (interview Q2).
  • Origins on a real surface. The same paper, Figure 3, three premature-beat origins estimated on electroanatomic surfaces next to the clinically identified sites [Zhou et al. (2021); free full text]. This module: Equation 3.
  • From digital twins to a classifier. Doste and colleagues, Figure 1, the pipeline from patient-specific heart models to simulated ECGs to a classifier tested on clinical ECGs, and Figure 2, simulated and clinical V2 and V3 QRS complexes side by side [Doste et al. (2022); free full text]. Compare with lab part E.

Further reading

  • Regression and patient-specific localization. Zhou, AbdelWahab, et al. (2018), Zhou, Horáček, et al. (2018), Zhou et al. (2020), Zhou et al. (2019), Zhou et al. (2021).
  • Active and hybrid methods. Missel et al. (2020), Meisenzahl et al. (2024); a tutorial on Bayesian optimization by Frazier (2018).
  • Simulation-trained models and data. Doste et al. (2022), Zheng et al. (2020).
  • Small cohorts and reviews. Botis et al. (2026), Yang et al. (2024), Antoun et al. (2026).
  • AI competency areas. Machine learning; Bayesian networks (Gaussian processes); search (shortest paths on the mesh); Markov decision processes (choosing the next pacing site). See the AI competency map.

References

Antoun, Ibrahim, Alkassem Alkhayer, Ahmed Abdelrazik, et al. 2026. “Artificial Intelligence Enabled Noninvasive Mapping of Cardiac Arrhythmia Origins Using the 12-Lead Electrocardiogram.” Journal of Cardiovascular Electrophysiology 37 (4): 847–53. https://doi.org/10.1111/jce.70319.
Botis, Michail, Dimitris Tsiachris, George G Botis, et al. 2026. “Machine Learning-Based Classification of Right and Left Ventricular Outflow Tract Premature Ventricular Contractions.” Annals of Noninvasive Electrocardiology 31 (6): e70250. https://doi.org/10.1111/anec.70250.
Doste, Ruben, Miguel Lozano, Guillermo Jimenez-Perez, et al. 2022. “Training Machine Learning Models with Synthetic Data Improves the Prediction of Ventricular Origin in Outflow Tract Ventricular Arrhythmias.” Frontiers in Physiology 13: 909372. https://doi.org/10.3389/fphys.2022.909372.
Frazier, Peter I. 2018. A Tutorial on Bayesian Optimization. arXiv:1807.02811. https://doi.org/10.48550/arXiv.1807.02811.
Meisenzahl, Casey, Karli Gillette, Anton J Prassl, et al. 2024. “BOATMAP: Bayesian Optimization Active Targeting for Monomorphic Arrhythmia Pace-Mapping.” Computers in Biology and Medicine 182: 109201. https://doi.org/10.1016/j.compbiomed.2024.109201.
Missel, Ryan, Prashnna K Gyawali, Jaideep Vitthal Murkute, et al. 2020. “A Hybrid Machine Learning Approach to Localizing the Origin of Ventricular Tachycardia Using 12-Lead Electrocardiograms.” Computers in Biology and Medicine 126: 104013. https://doi.org/10.1016/j.compbiomed.2020.104013.
Yang, Rui, Yiwen Wang, Yanan Wang, et al. 2024. “Machine Learning for Localization of Premature Ventricular Contraction Origins: A Review.” Pacing and Clinical Electrophysiology 47 (11): 1481–91. https://doi.org/10.1111/pace.15089.
Zheng, Jianwei, Guohua Fu, Kyle Anderson, et al. 2020. “A 12-Lead ECG Database to Identify Origins of Idiopathic Ventricular Arrhythmia Containing 334 Patients.” Scientific Data 7 (1): 98. https://doi.org/10.1038/s41597-020-0440-8.
Zhou, Shijie, Amir AbdelWahab, B Milan Horáček, et al. 2020. “Prospective Assessment of an Automated Intraprocedural 12-Lead ECG-Based System for Localization of Early Left Ventricular Activation.” Circulation. Arrhythmia and Electrophysiology 13 (7): e008262. https://doi.org/10.1161/CIRCEP.119.008262.
Zhou, Shijie, Amir AbdelWahab, John L Sapp, et al. 2018. “Localization of Ventricular Activation Origin from the 12-Lead ECG: A Comparison of Linear Regression with Non-Linear Methods of Machine Learning.” Annals of Biomedical Engineering 47 (2): 403–12. https://doi.org/10.1007/s10439-018-02168-y.
Zhou, Shijie, Amir AbdelWahab, John L Sapp, et al. 2021. “Assessment of an ECG-Based System for Localizing Ventricular Arrhythmias in Patients with Structural Heart Disease.” Journal of the American Heart Association 10 (20): e022217. https://doi.org/10.1161/JAHA.121.022217.
Zhou, Shijie, B Milan Horáček, James W Warren, et al. 2018. “Rapid 12-Lead Automated Localization Method: Comparison to Electrocardiographic Imaging (ECGI) in Patient-Specific Geometry.” Journal of Electrocardiology 51 (6S): S92–97. https://doi.org/10.1016/j.jelectrocard.2018.07.022.
Zhou, Shijie, John L Sapp, B Milan Horáček, et al. 2019. “Automated Intraprocedural Localization of Origin of Ventricular Activation Using Patient-Specific Computed Tomographic Imaging.” Heart Rhythm 17 (4): 567–75. https://doi.org/10.1016/j.hrthm.2019.10.025.