Skip to main content

m25 · Conduction velocity estimation: triangulation, surface fitting, omnipolar methods

Warning

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

In 60 seconds

Conduction velocity is the inverse of the activation-time gradient: \(\mathbf v = \nabla T/\lvert\nabla T\rvert^2\). Every estimator differs in how it gets \(\nabla T\), and all of them inherit timing error. For two points \(h\) apart, the relative speed error is \(\sqrt2\,\sigma c/h\), so fast conduction and close points make speed noisy; 1 ms of error at 1 m/s over 4 mm is 35%. Triangulation uses three points and is as good as the pair for an equilateral triangle, but much worse for a thin one. A local plane fit to the \(n\) points within radius \(R\) cuts the noise to about \(2\sigma/(R\sqrt n)\) per component, at the cost of bias where activation curves. Because speed is an inverse, noise in the gradient also biases speed low. Potential-based (omnipolar) methods skip activation times: for a plane wave, \(-\phi_t\nabla\phi/\lvert\nabla\phi\rvert^2\) is exact for any waveform, but a finite clique of electrodes biases it, by an amount that grows with spacing over electrode height.

From the field

  • Fitting surfaces to activation times. Bayly and colleagues fitted local polynomial surfaces \(T(x, y)\) to activation times and took speed and direction from their gradient. They validated this on simulations, then used it on pig epicardial recordings from arrays of over 500 electrodes (Bayly et al. 1998).
  • Speed differs with tissue voltage. In one left-atrial study, conduction velocity averaged about 1.0 m/s in normal-voltage areas of the anterior wall and 0.6 m/s in low-voltage areas (Heil et al. 2023).
  • Velocity vectors in tachycardia maps. In 20 patients with complex atrial tachycardias, a mapping method that built velocity vectors from delays between neighboring sites identified isthmus or breakout sites where ablation terminated the tachycardia in 25 of 26 tachycardias (96.2%), against 69.2% for conventional maps, as the study reports (Vicera et al. 2020).

Derivations

Symbols used throughout. \(T\) activation time; \(\mathbf g = \nabla T\); \(\mathbf v\) velocity; \(c\) speed; \(\sigma\) timing SD; \(\sigma_g\) gradient SD per component; \(h\) point spacing; \(E\) edge matrix of a triangle; \(S\) covariance pattern of time differences; \(X_c\) centered coordinates; \(n\) number of points; \(R\) fit radius; \(\phi\) potential; \(\phi_t\) its time derivative; \(U\), \(a\) the m03 waveform and its length scale; \(d\) clique spacing.

Speed from a gradient

\[ \mathbf v = \frac{\nabla T}{\lvert\nabla T\rvert^2}, \qquad c = \frac{1}{\lvert\nabla T\rvert}, \qquad \mathrm E\lvert\hat{\mathbf g}\rvert^2 = \lvert\mathbf g\rvert^2 + 2\sigma_g^2 \text{ (2D)} \tag{1}\]

Noise inflates the gradient on average, so speeds read low, and because speed is an inverse, the spread grows with speed. With an illustrative gradient noise of 0.2 ms/mm, a true 1.0 m/s read 0.78–1.30 m/s (10th–90th percentile) and a true 2.0 m/s read 1.27–3.35 m/s (lab part A).

Two points and triangles

\[ \frac{\sigma_c}{c} = \frac{\sqrt2\,\sigma\,c}{h} \tag{2}\]

\[ E = \begin{pmatrix} (\mathbf x_2 - \mathbf x_1)^\top \\ (\mathbf x_3 - \mathbf x_1)^\top \end{pmatrix}, \qquad E\,\mathbf g = \begin{pmatrix} T_2 - T_1 \\ T_3 - T_1 \end{pmatrix}, \qquad \operatorname{Cov}(\hat{\mathbf g}) = \sigma^2 E^{-1} S E^{-\top}, \quad S = \begin{pmatrix} 2 & 1 \\ 1 & 2 \end{pmatrix} \tag{3}\]

Both follow by first-order error propagation. With 1 ms of timing error at 1.0 m/s, points 2 mm apart gave 71% relative error and 8 mm apart 18%. An equilateral triangle of side \(h\) has isotropic gradient SD \(\sqrt2\,\sigma/h\), the same as the pair; a 4-mm sliver 0.4 mm high had 3.06 ms/mm across its short side against 0.35 ms/mm (lab part B). Triangulation is one of the standard methods reviewed by Cantwell and colleagues (Cantwell et al. 2015).

Exact

Assumptions

  • Model: independent timing errors of equal SD; first-order propagation for two points or an equilateral triangle (Equation 2, Equation 3) and for a least-squares plane over a uniformly filled disk (Equation 4). Tested against epml.velocity.
  • Parameters: default speed 1.0 m/s (Heil et al. 2023); timing error and number of points are illustrative and adjustable.
  • What it ignores: bias from curved activation (lab part C), position errors (m16–m19), and the breakdown of the first-order formulas when errors are large.

Local plane fits

\[ \hat{\mathbf g} = \arg\min_{\mathbf g, t_0}\sum_i\bigl(T_i - t_0 - \mathbf g\cdot\mathbf x_i\bigr)^2, \qquad \operatorname{Cov}(\hat{\mathbf g}) = \sigma^2(X_c^\top X_c)^{-1} \approx \Bigl(\frac{2\sigma}{R\sqrt n}\Bigr)^2 I \text{ for a disk} \tag{4}\]

More points and a larger radius reduce noise, but a large radius averages over curved activation. 10 mm from a focal source, with dense illustrative sampling and 1 ms of error, the RMS speed error was 0.23 m/s at \(R\) = 2.5 mm, 0.04 m/s at 6 mm and 0.08 m/s at 9 mm, where bias dominated (lab part C). Bayly and colleagues fitted polynomial surfaces, which can follow some of that curvature (Bayly et al. 1998).

Potential-based (omnipolar) velocity

For a plane wave \(\phi(\mathbf x, t) = U(\hat n\cdot\mathbf x - ct)\) of any waveform \(U\):

\[ \mathbf v = -\frac{\phi_t\,\nabla\phi}{\lvert\nabla\phi\rvert^2}, \qquad \frac{\hat c}{c} = 1 + \frac{(d/2)^2}{a^2} \text{ for a cross-shaped clique at the front's arrival along an axis} \tag{5}\]

The first result needs no activation times. It is the idea behind the m03 clique velocity (eq-m03-clique-velocity) and behind omnipolar methods, as described by Massé, Deno, Magtibay and colleagues (Massé et al. 2016; Deno et al. 2017; Magtibay et al. 2019). In practice \(\nabla\phi\) comes from differences between electrodes. For the cross-shaped clique here, with the m03 waveform, a 2-mm clique at 1 mm height read speed 1.59 times too fast. At 40° to the axis it read 1.30 times, with the direction biased toward the diagonal (lab part D). Other stencils and time windows have other biases.

Exact

Assumptions

  • Model: the m03 plane wave; a center electrode for \(\phi_t\) and four electrodes at \(\pm d/2\) along two axes for the gradient by central differences, evaluated as the front reaches the center (Equation 5). Tested against epml.velocity.
  • Parameters: speed 1.0 m/s (Heil et al. 2023); front width 0.3 mm (m03); height 1 mm by default (Riccio et al. 2021); spacing and angle adjustable.
  • What it ignores: the field-loop and fitting steps of published omnipolar methods, other clique shapes, curved fronts and noise.

Key takeaways

  • Speed is noisy where it is fast and where points are close; average gradients, not speeds.
  • Every estimator trades noise for bias through its spacing or radius.

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 this shows up in EP. Conduction velocity maps are used to look for slow conduction, which tends to occur in low-voltage tissue (Heil et al. 2023). Velocity vectors from neighboring sites have been used in atrial tachycardia maps (Vicera et al. 2020). Omnipolar methods estimate velocity from potentials in a grid of electrodes (Magtibay et al. 2019). Every such estimate depends on timing error, point spacing and the estimator (Equation 2, Equation 4).

What the evidence is. A methods paper with animal data (Bayly et al. 1998), a review of algorithms (Cantwell et al. 2015), an observational clinical series (Vicera et al. 2020), a clinical study of regional velocities (Heil et al. 2023) and a review of omnipolar methods (Magtibay et al. 2019).

What a reader should ask:

Visual summary

Schematic

triangle plane fit clique
Figure 1: Three ways to estimate velocity: from a triangle of points, a local plane fit, or a clique’s potentials.

Assumptions

How experts think

Before trusting a conduction velocity, ask:

  1. How big is the timing error compared with the travel time between points? (Equation 2)
  2. What shape are the triangles? (Equation 3)
  3. How large is the fit radius compared with the curvature of activation? (Equation 4)
  4. Was the speed averaged as an inverse of noisy gradients? (Equation 1)
  5. For potential-based velocity, how big is the clique compared with the electrode height? (Equation 5)

One sentence to keep: speed is an inverse; noise makes fast conduction look erratic and slow regions look slow.

Cheat sheet

Quantity Result
Gradient (Equation 1) \(\mathbf v = \nabla T/\lvert\nabla T\rvert^2\); \(\mathrm E\lvert\hat g\rvert^2 = \lvert g\rvert^2 + 2\sigma_g^2\)
Two points (Equation 2) \(\sigma_c/c = \sqrt2\,\sigma c/h\)
Triangle (Equation 3) \(E\mathbf g = \Delta T\); \(\operatorname{Cov} = \sigma^2E^{-1}SE^{-\top}\)
Plane fit (Equation 4) \(\operatorname{Cov} = \sigma^2(X_c^\top X_c)^{-1} \approx (2\sigma/(R\sqrt n))^2 I\)
Potential (Equation 5) \(\mathbf v = -\phi_t\nabla\phi/\lvert\nabla\phi\rvert^2\); cross clique \(\hat c/c = 1 + (d/2)^2/a^2\)

Values used in this module

Value Source
About 1.0 m/s in normal-voltage and 0.6 m/s in low-voltage left atrium Heil et al. (2023)
Isthmus or breakout sites identified in 96.2% vs 69.2% of 26 tachycardias (20 patients) Vicera et al. (2020)
Polynomial surface fitting; 500-plus-electrode arrays Bayly et al. (1998)
Timing and gradient errors, point counts, radii, geometry illustrative
Relative errors, covariances, RMS errors, clique ratios our own code

Interview questions

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

Q1. How do you get conduction velocity from an activation map?

Model answer

\(\mathbf v = \nabla T/\lvert\nabla T\rvert^2\); speed is the inverse of the gradient magnitude (Equation 1).

Q2. Why does noise bias speed low?

Model answer

Noise adds to the gradient’s magnitude on average, \(\mathrm E\lvert\hat g\rvert^2 = \lvert g\rvert^2 + 2\sigma_g^2\), and speed is its inverse (Equation 1).

Q3. Two points 4 mm apart, 1 ms timing error each, 1 m/s. Relative speed error?

Model answer

\(\sqrt2 \times 1 \times 1/4 \approx\) 35% (Equation 2).

Q4. Why is fast conduction harder to measure?

Model answer

The travel time between points is short compared with the timing error; relative error grows in proportion to \(c\) (Equation 2).

Q5. How do you compute a gradient from three points, and when is it unreliable?

Model answer

Solve \(E\mathbf g = \Delta T\); the covariance \(\sigma^2E^{-1}SE^{-\top}\) blows up for thin triangles (Equation 3) (Cantwell et al. 2015).

Q6. What does a local plane fit buy you?

Model answer

Averaging: gradient SD about \(2\sigma/(R\sqrt n)\); but a large radius biases the estimate where activation curves (Equation 4) (Bayly et al. 1998).

Q7. How would you choose the fit radius?

Model answer

Balance noise, which falls with radius, against curvature bias, which grows; test on simulated data with known speed (Equation 4).

Q8. How can a clique of electrodes estimate velocity without activation times?

Model answer

For a plane wave, \(\mathbf v = -\phi_t\nabla\phi/\lvert\nabla\phi\rvert^2\) for any waveform; a clique supplies \(\nabla\phi\) and \(\phi_t\) (Equation 5) (Massé et al. 2016; Deno et al. 2017).

Q9. What limits the clique estimate?

Model answer

Finite differences over spacing \(d\) misestimate \(\nabla\phi\) by order \((d/a)^2\); curved fronts and noise add more (Equation 5).

Q10. You average speeds over a region with noisy gradients. What goes wrong?

Model answer

Speed is an inverse; a few small noisy gradients give huge speeds that dominate the mean. Average gradients, or use medians (Equation 1).

Q11. What speeds were reported for normal- and low-voltage left atrium in one study?

Model answer

About 1.0 m/s and 0.6 m/s (Heil et al. 2023).

Q12. Name three families of conduction velocity estimators.

Model answer

Triangulation and finite differences of activation times, local surface fitting, and potential-based (omnipolar) methods (Equation 3, Equation 4, Equation 5) (Cantwell et al. 2015).

Code lab

The lab estimates conduction velocity from synthetic activation.

Part What you do Data
A Speed from a gradient (Equation 1) synthetic
B Two points and triangles (Equation 2, Equation 3) synthetic
C Local plane fits (Equation 4) synthetic
D Velocity from a clique (Equation 5) synthetic
E A conduction velocity map 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. Gradient 2 ms/mm. Speed?

Answer

0.5 mm/ms = 0.5 m/s (Equation 1).

S2. Double the spacing between two points. Relative speed error?

Answer

Halved (Equation 2).

S3. Same spacing, speed doubles. Relative error?

Answer

Doubled (Equation 2).

S4. A plane fit uses 4 times as many points over the same disk. Gradient SD?

Answer

Halved (Equation 4).

S5. Cross clique, d = 2 mm, a = 2 mm, along an axis. Speed ratio?

Answer

\(1 + 1/4 =\) 1.25 (Equation 5).

S6. Gradient 1 ms/mm with noise SD 0.5 ms/mm per component. RMS gradient magnitude?

Answer

\(\sqrt{1 + 0.5} \approx\) 1.22 ms/mm (Equation 1).

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.

  • Velocity vector fields. Bayly and colleagues’ vector fields from polynomial surface fits during pacing and fibrillation (Bayly et al. 1998). This module: local plane fits.
  • Algorithm comparison. Cantwell and colleagues’ overview of annotation and velocity methods (Cantwell et al. 2015). This module: two points and triangles.
  • Omnipolar velocity. Magtibay and colleagues’ illustrations of omnipolar activation vectors (Magtibay et al. 2019). This module: potential-based velocity.

Further reading

References

Bayly, P. V., B. H. KenKnight, J. M. Rogers, R. E. Hillsley, R. E. Ideker, and W. M. Smith. 1998. “Estimation of Conduction Velocity Vector Fields from Epicardial Mapping Data.” IEEE Transactions on Biomedical Engineering 45 (5): 563–71. https://doi.org/10.1109/10.668746.
Cantwell, C D, C H Roney, F S Ng, et al. 2015. “Techniques for Automated Local Activation Time Annotation and Conduction Velocity Estimation in Cardiac Mapping.” Computers in Biology and Medicine 65: 229–42. https://doi.org/10.1016/j.compbiomed.2015.04.027.
Deno, D. Curtis, Ram Balachandran, Dennis Morgan, Faisal Ahmad, et al. 2017. “Orientation-Independent Catheter-Based Characterization of Myocardial Activation.” IEEE Transactions on Biomedical Engineering 64 (5): 1067–77. https://doi.org/10.1109/TBME.2016.2589158.
Heil, Emanuel, Jin-Hong Gerds-Li, Julian Keznickl-Pulst, Vesna Furundzija-Cabraja, et al. 2023. “Left Atrial Conduction Times and Regional Velocities in Persistent Atrial Fibrillation Patients with and Without Fibrotic Atrial Cardiomyopathy.” Heart and Vessels 38 (10): 1277–87. https://doi.org/10.1007/s00380-023-02282-5.
Magtibay, Karl, Andreu Porta-Sánchez, Shouvik K. Haldar, et al. 2019. “Reinserting Physiology into Cardiac Mapping Using Omnipolar Electrograms.” Cardiac Electrophysiology Clinics 11 (3): 525–36. https://doi.org/10.1016/j.ccep.2019.05.003.
Massé, Stéphane, Karl Magtibay, Nicholas Jackson, John Asta, et al. 2016. “Resolving Myocardial Activation with Novel Omnipolar Electrograms.” Circulation: Arrhythmia and Electrophysiology 9 (7): e004107. https://doi.org/10.1161/CIRCEP.116.004107.
Riccio, Jennifer, Alejandro Alcaine, Sara Rocher, Laura Martinez-Mateu, et al. 2021. “Characterization of Atrial Propagation Patterns and Fibrotic Substrate with a Modified Omnipolar Electrogram Strategy in Multi-Electrode Arrays.” Frontiers in Physiology 12: 674223. https://doi.org/10.3389/fphys.2021.674223.
Vicera, Jennifer Jeanne B., Yenn-Jiang Lin, Po-Tseng Lee, et al. 2020. “Identification of Critical Isthmus Using Coherent Mapping in Patients with Scar-Related Atrial Tachycardia.” Journal of Cardiovascular Electrophysiology 31 (6): 1436–47. https://doi.org/10.1111/jce.14457.