Skip to main content

m26 · Map quality and uncertainty: point density, annotation error, interpolation error

Warning

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

In 60 seconds

A map is only as good as its points and the gaps between them, and a good map says where it is unsure. A Gaussian process (GP) interpolates activation times and gives, at every site, a mean and a variance. The variance is small near points and grows in gaps. Without ground truth, leave-one-out error scores a map and helps choose its smoothness, and for a GP it needs no refitting. Each point brings its own error: annotation, reference (if points come from different beats) and position, which moves a time by distance over conduction speed. These add in variance to give the GP’s observation noise. Point density sets the typical distance to the nearest point, about \(1/(2\sqrt\rho)\): 5 mm at 1 point/cm². Because the GP variance depends only on where the points are, the next point can be planned where the map is least sure.

From the field

  • Enough points depends on the rhythm. In simulations and 21 clinical atrial tachycardia maps, the density beyond which re-interpolated maps improved little was about 0.67 points/cm² for focal activation, 1.05 for macroreentry and 1.23 for localized reentry in the left atrium; about 1.0–1.5 points/cm² overall (Williams et al. 2018).
  • Probabilistic maps on the atrium. Coveney and colleagues interpolated uncertain activation times with Gaussian processes directly on atrial surfaces. They found that conduction velocity uncertainty depends on data density, propagation direction and speed (Coveney et al. 2020).
  • Errors come from every stage. Annotation and reference choices (m22), position errors (m16–m19) and interpolation (m24) each contribute; open tools such as OpenEP let researchers recompute maps from exported data (Williams et al. 2021).

Derivations

Symbols used throughout. \(\mathbf x\) site; \(\mathbf y\) measured times at sites \(X\); \(m\) prior mean; \(k\) covariance (kernel) with variance \(s^2\) and length scale \(\ell\); \(K\) kernel matrix; \(\mathbf k_*\) covariances with a new site; \(\sigma_n\) per-point error SD; \(A = K + \sigma_n^2 I\); \(r_i\) leave-one-out residual; \(\sigma_a\), \(\sigma_r\), \(\sigma_x\) annotation, reference and position SDs; \(c\) conduction speed; \(\rho\) point density; \(D\) distance to the nearest point.

A probabilistic map

With a squared-exponential kernel \(k(\mathbf x, \mathbf x') = s^2\exp(-\lvert\mathbf x - \mathbf x'\rvert^2/2\ell^2)\):

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

This is standard GP regression (Rasmussen and Williams 2006), which Coveney and colleagues adapted to atrial surfaces (Coveney et al. 2020). In the lab, 40 points with 1.5 ms of error mapped a focal activation with an RMS error of 1.9 ms at \(\ell\) = 8 mm, with SD ranging from below 1.5 ms at points to 16 ms in gaps. A short length scale gave wide, honest but uninformative bands; a long one gave tighter bands that held the truth at 94% of sites (lab part A).

Simulation

Assumptions

  • Model: GP regression in one dimension with a squared-exponential kernel (prior SD 20 ms) and a constant prior mean equal to the mean of the data (Equation 1); random point positions and Gaussian errors with a fixed seed. Tested against epml.uncertainty.
  • Parameters: focal activation at 1.0 m/s (Heil et al. 2023), 5 mm off the line; number of points, error and length scale are illustrative and adjustable.
  • What it ignores: fitting the kernel to data (part B of the lab), surfaces and geodesic distances, and correlated errors.

Leave-one-out without refitting

\[ r_i = y_i - \mu_{-i}(\mathbf x_i) = \frac{\bigl[A^{-1}(\mathbf y - m)\bigr]_i}{\bigl[A^{-1}\bigr]_{ii}}, \qquad \sigma^2_{-i} = \frac{1}{\bigl[A^{-1}\bigr]_{ii}} \tag{2}\]

One matrix inverse gives every leave-one-out prediction (Rasmussen and Williams 2006). In the lab, the closed form matched refitting to \(10^{-13}\) ms. The length scale with the lowest leave-one-out error (16 mm, 1.24 ms) was also the one with the lowest true error (1.46 ms) (lab part B).

The per-point error budget

\[ \sigma_n^2 = \sigma_a^2 + \sigma_r^2 + \Bigl(\frac{\sigma_x}{c}\Bigr)^2 \tag{3}\]

Independent errors add in variance. A position error \(\sigma_x\) along the propagation direction shifts the time by \(\sigma_x/c\); at 1.0 m/s (Heil et al. 2023), 2 mm becomes 2 ms. The reference term applies when points come from different beats (m22, eq-m22-lat). With an illustrative 1 ms of annotation, 2 ms of reference and 2 mm of position error, the per-point SD was 3.0 ms, and the reference term was the largest (lab part C).

Point density

For points placed at random with density \(\rho\), the distance \(D\) from a random site to the nearest point satisfies \(P(D > r) = e^{-\rho\pi r^2}\):

\[ \mathrm E[D] = \frac{1}{2\sqrt\rho}, \qquad \operatorname{median}(D) = \sqrt{\frac{\ln 2}{\pi\rho}} \tag{4}\]

At 1.0 points/cm² the mean is 5.0 mm and at 1.5 points/cm² 4.1 mm, the range Williams and colleagues found sufficient for atrial tachycardia maps (Williams et al. 2018). The same law gives the m24 coverage fraction.

Exact

Assumptions

  • Model: the per-point budget of Equation 3; the GP SD at a given distance from a single point with that observation noise (Equation 1); the mean nearest-point distance of Equation 4. Tested against epml.uncertainty.
  • Parameters: speed 1.0 m/s (Heil et al. 2023); kernel length scale 8 mm and prior SD 20 ms are illustrative; error sizes and density are adjustable; the default density is within the 1.0–1.5 points/cm² range (Williams et al. 2018).
  • What it ignores: neighboring points beyond the nearest, biases (as opposed to random errors), and errors that are shared between points.

Where to map next

\[ \mathbf x_{n+1} = \arg\max_{\mathbf x}\,\sigma_n^2(\mathbf x) \tag{5}\]

The posterior variance of Equation 1 does not depend on the measured values, only on where they were taken, so a sampling plan can be made before mapping. In the lab, 40 points placed one at a time where the variance was largest cut the largest SD from 12.0 ms (random placement) to 3.5 ms and the RMS error from 2.0 to 1.2 ms (lab part E). Real mapping is constrained by catheter reach and contact.

Key takeaways

  • Report a map’s uncertainty with the map; it is largest in the gaps.
  • Errors add in variance; position error counts in time as distance over speed.

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. How many points a map needs depends on the rhythm and chamber (Williams et al. 2018). Probabilistic interpolation has been demonstrated on clinical atrial data, with conduction velocity uncertainty that depends on data density and propagation (Coveney et al. 2020). Maps are often shown without uncertainty, so a reader has to judge it from point density, annotation and the other error sources in this part of the course.

What the evidence is. A combined simulation and clinical study of sampling density (Williams et al. 2018), a methods study with two clinical cases and simulated validation (Coveney et al. 2020), and GP theory (Rasmussen and Williams 2006).

What a reader should ask:

  • Where are the points, and how far is this site from the nearest one? (Equation 4)
  • What error does each point carry? (Equation 3)
  • How smooth was the map assumed to be, and was that checked? (Equation 1, Equation 2)
  • Would one more point here change the conclusion? (Equation 5)

Visual summary

Schematic

uncertainty grows in gaps errors add in variance map next where least sure
Figure 1: Uncertainty grows between points, per-point errors add in variance, and the next point goes where the map is least sure.

Assumptions

How experts think

Before trusting a feature on a map, ask:

  1. How many points support it, and how far away are they? (Equation 4)
  2. How large is each point’s error, and which source dominates? (Equation 3)
  3. Is the feature larger than the map’s uncertainty there? (Equation 1)
  4. Does it survive leaving out single points? (Equation 2)
  5. Where would one more point help most? (Equation 5)

One sentence to keep: a map without uncertainty hides where it was guessed.

Cheat sheet

Quantity Result
GP (Equation 1) \(\mu = m + \mathbf k_*^\top A^{-1}(\mathbf y - m)\); \(\sigma^2 = s^2 - \mathbf k_*^\top A^{-1}\mathbf k_*\)
Leave-one-out (Equation 2) \(r_i = [A^{-1}(\mathbf y - m)]_i/[A^{-1}]_{ii}\)
Budget (Equation 3) \(\sigma_n^2 = \sigma_a^2 + \sigma_r^2 + (\sigma_x/c)^2\)
Density (Equation 4) \(\mathrm E[D] = 1/(2\sqrt\rho)\); median \(\sqrt{\ln 2/(\pi\rho)}\)
Next point (Equation 5) \(\arg\max\sigma_n^2(\mathbf x)\); variance is independent of the measured values

Values used in this module

Value Source
About 1.0–1.5 points/cm² overall; 0.67, 1.05 and 1.23 by mechanism Williams et al. (2018)
CV uncertainty depends on density, direction and speed Coveney et al. (2020)
Conduction velocity 1.0 m/s Heil et al. (2023)
Kernel settings, error sizes, point counts illustrative
RMS errors, coverages, distances, SDs 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. What does a Gaussian process add to interpolation?

Model answer

A variance at every site, small near points and large in gaps (Equation 1) (Rasmussen and Williams 2006).

Q2. How would you check that a map’s uncertainty is honest?

Model answer

On data with known truth, count how often the truth lies within the stated interval; on real data, compare leave-one-out residuals with their predicted variances (Equation 1, Equation 2).

Q3. How do you get leave-one-out errors for a GP cheaply?

Model answer

From one inverse: \(r_i = [A^{-1}(\mathbf y - m)]_i/[A^{-1}]_{ii}\) (Equation 2) (Rasmussen and Williams 2006).

Q4. What does the kernel length scale control?

Model answer

How far each point’s influence reaches; too short leaves wide bands, too long oversmooths. Choose it by leave-one-out or likelihood (Equation 1, Equation 2).

Q5. List the errors one mapped point carries.

Model answer

Annotation, reference (when points come from different beats) and position; independent ones add in variance (Equation 3).

Q6. How does a position error become a time error?

Model answer

Along the propagation direction it shifts the time by distance over speed: 2 mm at 1 m/s is 2 ms (Equation 3).

Q7. Points at 1 per cm². Typical distance from a site to the nearest point?

Model answer

About 5 mm on average for random placement (Equation 4).

Q8. How many points did one study find sufficient for atrial tachycardia maps?

Model answer

About 1.0–1.5 points/cm², more for reentry than for focal activation (Williams et al. 2018).

Q9. Why can you plan the next mapping point before measuring?

Model answer

A GP’s variance depends only on point locations, not values (Equation 5).

Q10. Greedy maximum-variance sampling: what is it good and bad at?

Model answer

Good at filling gaps fast; it ignores catheter reach and contact, and it trusts the kernel, so a wrong length scale misplaces points (Equation 5).

Q11. What did Coveney and colleagues find drives conduction velocity uncertainty?

Model answer

Data density, propagation direction and speed (Coveney et al. 2020).

Q12. A feature appears in a gap between points. What do you check?

Model answer

Whether it is larger than the map’s uncertainty there, and whether one more point would confirm it (Equation 1, Equation 5).

Code lab

The lab measures map quality on synthetic data.

Part What you do Data
A A probabilistic activation map (Equation 1) synthetic
B Leave-one-out without refitting (Equation 2) synthetic
C The per-point error budget (Equation 3) synthetic
D Point density (Equation 4) synthetic
E Where to map next (Equation 5) 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. Annotation 1 ms, reference 2 ms, position 2 mm at 1 m/s. Per-point SD?

Answer

\(\sqrt{1 + 4 + 4} =\) 3 ms (Equation 3).

S2. The same points all come from one beat. Per-point SD for differences between them?

Answer

The reference term drops: \(\sqrt{1 + 4} \approx\) 2.2 ms per point (Equation 3).

S3. Density quadruples. Mean nearest-point distance?

Answer

Halves (Equation 4).

S4. Far from all points, what is the GP’s SD?

Answer

The prior SD \(s\) (Equation 1).

S5. At a point measured with error SD 2 ms and a very broad prior, the GP’s SD is about what?

Answer

2 ms (Equation 1).

S6. You change the measured values but not the locations. Does the GP’s variance change?

Answer

No (Equation 5).

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.

  • Error against sampling density. Williams and colleagues’ curves of map error against points per cm² (Williams et al. 2018). This module: point density.
  • Probabilistic atrial maps. Coveney and colleagues’ mean and SD maps of activation and conduction velocity on atrial surfaces (Coveney et al. 2020). This module: a probabilistic map.
  • GP regression. Rasmussen and Williams’ illustrations of GP posteriors and length scales (Rasmussen and Williams 2006). This module: a probabilistic map and leave-one-out.

Further reading

  • Theory. Gaussian process regression and cross-validation (Rasmussen and Williams 2006).
  • Cardiac application. Probabilistic interpolation on atrial surfaces (Coveney et al. 2020); sampling density (Williams et al. 2018); open analysis of mapping data (Williams et al. 2021).
  • Related modules. m16–m19 (position errors), m22 (annotation), m24 (interpolation), m25 (conduction velocity), m30 (capstone), m46 (trustworthy AI and calibration).

References

Coveney, Sam, Cesare Corrado, Caroline H. Roney, et al. 2020. “Gaussian Process Manifold Interpolation for Probabilistic Atrial Activation Maps and Uncertain Conduction Velocity.” Philosophical Transactions of the Royal Society A 378 (2173): 20190345. https://doi.org/10.1098/rsta.2019.0345.
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.
Rasmussen, Carl Edward, and Christopher K. I. Williams. 2006. Gaussian Processes for Machine Learning. Adaptive Computation and Machine Learning. MIT Press. https://doi.org/10.7551/mitpress/3206.001.0001.
Williams, Steven E. et al. 2021. “OpenEP: A Cross-Platform Electroanatomic Mapping Data Format and Analysis Platform for Electrophysiology Research.” Frontiers in Physiology 12: 646023. https://doi.org/10.3389/fphys.2021.646023.
Williams, Steven E., James L. Harrison, Henry Chubb, et al. 2018. “Local Activation Time Sampling Density for Atrial Tachycardia Contact Mapping: How Much Is Enough?” Europace 20 (2): e11–20. https://doi.org/10.1093/europace/eux037.