m18 · Hybrid and dielectric localization concepts
Education only. Not for diagnosis, treatment or any clinical decision.
In 60 seconds
Magnetic localization is metric but needs sensor coils; impedance localization sees every electrode but is distorted (m16, m17). A hybrid system uses positions where both are known, from a catheter that carries a magnetic sensor and electrodes, to learn a map from impedance coordinates to true ones, and applies it to every other electrode. An affine map fixes scale and shear; a smooth correction, here Gaussian radial basis functions plus an affine part, also removes the nonlinearity, if reference points cover the region. Where two estimates exist, weighting them by inverse variance gives a combination at least as precise as the better one. Dielectric approaches go further: tissue both conducts and polarizes, \(\sigma^* = \sigma + j\omega\varepsilon_0\varepsilon_r(\omega)\), so impedance measured at several frequencies carries information about what lies around the electrodes, not just where they are.
From the field
- Dielectric imaging builds anatomy from fields. A review describes systems that use electrical fields to distinguish anatomic structures by their conductivity and permittivity, enabling non-fluoroscopic, noncontact mapping of anatomy (Abeln et al. 2021).
- Its anatomy has been compared with angiography. In 65 patients (260 pulmonary veins), a wide-band dielectric system overestimated angiographic ostial diameters by 1.08 mm on average, with 95% limits of agreement of −1.93 to 4.1 mm (Rottner et al. 2021).
- Impedance phase carries tissue information. In a pig infarct model (137 endocardial sites in 5 animals), infarct scar showed a 37.4% lower impedance magnitude and a 21.5% lower phase angle than healthy tissue; four impedance parameters together gave an area under the ROC curve of 0.96 (Amorós-Figueras et al. 2018).
Derivations
Symbols used throughout. \(\mathbf u\) impedance coordinates (mm, after the uniform-field map of m17); \(\mathbf x\) true coordinates; \(A\), \(\mathbf b\) affine map; \(\varphi\) a radial basis function of width \(s\); \(\mathbf w_j\) basis weights; \(\lambda\) regularization; \(h\) distance to the nearest reference point; \(f\) a smooth function and \(f''\) its second derivative; \(\hat{\mathbf x}_1\), \(\hat{\mathbf x}_2\) two estimates with variances \(v_1\), \(v_2\); \(\sigma\) conductivity; \(\varepsilon_0\) vacuum permittivity; \(\varepsilon_r\) relative permittivity; \(\omega = 2\pi f\); \(\tau\) relaxation time; \(\varepsilon_\infty\), \(\Delta\varepsilon\) high-frequency permittivity and dispersion strength.
An affine correction
At \(N\) reference points both \(\mathbf u_i\) and \(\mathbf x_i\) are known. The least-squares affine map is:
\[ (\hat A, \hat{\mathbf b}) = \arg\min_{A,\mathbf b}\sum_{i=1}^N \lVert \mathbf x_i - A\mathbf u_i - \mathbf b\rVert^2 \tag{1}\]
It removes per-axis scale errors and shear: in the lab it cut the error of a distorted impedance space from 3.2 mm RMS to about 1.35 mm, and no further, because the remaining error is nonlinear (lab part B).
A smooth correction
Add radial basis functions centered on the reference points:
\[ \mathbf x(\mathbf u) = A\mathbf u + \mathbf b + \sum_{j=1}^N \mathbf w_j\,\varphi(\lVert\mathbf u - \mathbf u_j\rVert), \qquad \varphi(r) = e^{-r^2/(2s^2)}, \qquad \begin{bmatrix}\Phi + \lambda I & P\\ P^\top & 0\end{bmatrix}\begin{bmatrix}W\\ C\end{bmatrix} = \begin{bmatrix}X\\ 0\end{bmatrix} \tag{2}\]
Here \(\Phi_{ij} = \varphi(\lVert\mathbf u_i - \mathbf u_j\rVert)\), \(P\) has rows \([\mathbf u_i^\top, 1]\), and \(C\) holds the affine part. With \(\lambda = 0\) it interpolates the reference points exactly; a small \(\lambda\) smooths noise. With 100 reference points spread through the chamber, the error fell to 0.07 mm; with only 10 it was worse than the affine map, because the correction overfit (lab part B). This is one generic way to build a hybrid correction; products do not publish theirs.
Simulation
Assumptions
- Model: the m17 point-patch impedance model with per-axis scale errors of 12%, −8% and 5% (illustrative); reference points drawn from a fixed random sequence inside a 45-mm disk; correction by Equation 2 with \(\lambda = 10^{-6}\), solved by Gaussian elimination. Tested against
epml.hybrid. - Parameters: the number of reference points and the basis width are adjustable.
- What it ignores: noise and drift in both systems, the third dimension, and changes of the field during the procedure.
Key takeaways
- Affine fixes scale; a smooth correction fixes shape, given enough well-spread reference points.
Coverage
Any correction is interpolation, and interpolation error grows with distance from the data. For linear interpolation on a segment of length \(h\):
\[ \lvert f(x) - f_{\text{lin}}(x)\rvert \le \frac{h^2}{8}\max\lvert f''\rvert \tag{3}\]
So the error scales as the square of the gap. In the lab, after a 25-point correction, the test points farther from their nearest reference point had three times the median error of the nearer half (lab part C). The same bound appears in m34 (eq-m34-interp-error).
Fusing two estimates
Where an electrode has two independent, unbiased estimates, the minimum-variance combination weights each by the inverse of its variance:
\[ \hat{\mathbf x} = \frac{\hat{\mathbf x}_1/v_1 + \hat{\mathbf x}_2/v_2}{1/v_1 + 1/v_2}, \qquad \operatorname{var}(\hat{\mathbf x}) = \frac{1}{1/v_1 + 1/v_2} \le \min(v_1, v_2) \tag{4}\]
With SDs of 0.5 and 1.5 mm, the fused SD is 0.47 mm (lab part D). The weights are only right if the variances are right, and the formula assumes no bias, which distortion violates (m16).
Key takeaways
- Corrections are interpolations: their error grows with distance from reference points.
- Inverse-variance weighting never does worse than the better estimate, if the variances are honest.
Dielectric properties
Tissue conducts and polarizes. Its complex conductivity, with one relaxation for illustration, is:
\[ \sigma^*(\omega) = \sigma + j\omega\varepsilon_0\varepsilon_r(\omega), \qquad \varepsilon_r(\omega) = \varepsilon_\infty + \frac{\Delta\varepsilon}{1 + j\omega\tau}, \qquad \arg Z = -\arg\sigma^* \tag{5}\]
Measured tissue spectra need several dispersion regions; Gabriel and colleagues fitted a parametric model from 10 Hz to 100 GHz for 17 tissue types (Gabriel et al. 1996). With two illustrative tissues, the impedance phases differed by 0.5° at 1 kHz and by 17° at 100 kHz (lab part E). Measuring at several frequencies separates materials that one frequency cannot, which is the idea behind dielectric approaches (Abeln et al. 2021).
Approximate model
Assumptions
- Model: Equation 5 with one Debye relaxation per tissue; \(\varepsilon_0\) from CODATA (Tiesinga et al. 2021). Tested against
epml.hybrid. - Parameters: both tissues’ conductivities, permittivities and relaxation times are illustrative, not measured.
- What it ignores: the several dispersion regions of real tissue (Gabriel et al. 1996), electrode polarization, geometry and contact.
Key takeaways
- Impedance is complex and frequency dependent; its phase carries tissue information.
Clinical correlate
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. Hybrid localization lets a mapping system show every electrode while keeping magnetic accuracy where it has a sensor (m16, m17). Dielectric imaging systems build anatomy from fields and have been compared with angiography for pulmonary vein dimensions (Abeln et al. 2021; Rottner et al. 2021). Local impedance magnitude and phase have been studied as markers of infarct scar in animals (Amorós-Figueras et al. 2018).
What the evidence is. A review (Abeln et al. 2021), a clinical comparison with angiography (Rottner et al. 2021), an animal study (Amorós-Figueras et al. 2018) and a tissue-property model (Gabriel et al. 1996).
What a reader should ask:
- Where were the reference points, and how far is this site from them? (Equation 3)
- Was the correction refreshed during the case? (m19)
- Which frequencies and parameters does a tissue claim rest on? (Equation 5)
Visual summary
Schematic
Assumptions
- Schematic: not to scale, no quantitative claim. Labels follow Equation 2 and Equation 5.
How experts think
When a system blends two kinds of localization, ask:
- Which positions are measured and which are inferred? (Equation 2)
- Where are the reference points? Error grows with the gap (Equation 3).
- Could the correction overfit? Few points and a flexible map are worse than an affine one (Equation 2).
- Are the variances used for weighting honest, and is either estimate biased? (Equation 4)
- For tissue claims: which frequencies, which parameters, validated against what? (Equation 5)
One sentence to keep: a hybrid correction is an interpolation; trust it near its reference points.
Cheat sheet
| Quantity | Result |
|---|---|
| Affine correction (Equation 1) | least squares for \(A\), \(\mathbf b\) |
| Smooth correction (Equation 2) | \(A\mathbf u + \mathbf b + \sum_j \mathbf w_j\varphi(\lVert\mathbf u - \mathbf u_j\rVert)\); block system with \(\lambda\) |
| Coverage (Equation 3) | linear interpolation error \(\le h^2\max\lvert f''\rvert/8\) |
| Fusion (Equation 4) | inverse-variance weights; variance \(1/(1/v_1 + 1/v_2)\) |
| Dielectric (Equation 5) | \(\sigma^* = \sigma + j\omega\varepsilon_0\varepsilon_r\); Debye \(\varepsilon_r = \varepsilon_\infty + \Delta\varepsilon/(1 + j\omega\tau)\) |
Values used in this module
| Value | Source |
|---|---|
| 65 patients, 260 veins; bias 1.08 mm; limits −1.93 to 4.1 mm | Rottner et al. (2021) |
| Scar: impedance magnitude −37.4%, phase −21.5%; AUC 0.96; 137 sites, 5 pigs | Amorós-Figueras et al. (2018) |
| Parametric model, 10 Hz–100 GHz, 17 tissue types | Gabriel et al. (1996) |
| \(\varepsilon_0\) | Tiesinga et al. (2021) |
| Scale errors, chamber size, SDs, tissue parameters | illustrative |
| Correction errors, fused SD, phases | 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. Why combine magnetic and impedance localization?
Model answer
Magnetic is metric but needs sensor coils; impedance sees every electrode but is distorted. Reference points with both let you correct the impedance space (Equation 1, Equation 2).
Q2. What does an affine correction fix, and what not?
Model answer
Scale, shear, rotation and offset; not the smooth nonlinearity (Equation 1).
Q3. How does a radial-basis correction work?
Model answer
An affine part plus weighted basis functions centered on the reference points, solved as one linear system; \(\lambda\) trades exact interpolation for smoothness (Equation 2).
Q4. With 10 reference points the smooth correction did worse than affine. Why?
Model answer
Too few points for a flexible map: it fits them and wanders between them (Equation 2, Equation 3).
Q5. Where is a correction least reliable?
Model answer
Far from reference points; interpolation error grows as the square of the gap (Equation 3).
Q6. Derive inverse-variance weighting.
Model answer
Minimize \(w^2v_1 + (1 - w)^2v_2\) over \(w\): \(w = v_2/(v_1 + v_2)\), giving variance \(1/(1/v_1 + 1/v_2)\) (Equation 4).
Q7. When does inverse-variance fusion mislead?
Model answer
When an estimate is biased or its variance is wrong; it assumes unbiased, independent estimates (Equation 4).
Q8. What is complex conductivity?
Model answer
\(\sigma + j\omega\varepsilon_0\varepsilon_r\): conduction plus displacement current; the second grows with frequency (Equation 5).
Q9. Why measure at several frequencies?
Model answer
Tissues relax at different frequencies, so their phases differ most near their relaxations; one frequency may not separate them (Equation 5) (Gabriel et al. 1996).
Q10. What does dielectric imaging use to distinguish structures?
Model answer
Differences in conductivity and permittivity, sensed through electrical fields (Abeln et al. 2021).
Q11. How was a dielectric system’s anatomy validated in one study?
Model answer
Pulmonary vein ostial diameters against angiography: a mean overestimate of 1.08 mm, limits of agreement −1.93 to 4.1 mm (Rottner et al. 2021).
Q12. An animal study found impedance phase lower in scar. What would you ask before using it in people?
Model answer
Contact and catheter dependence, frequency, reproducibility, and validation in patients with independent scar definitions; 5 pigs is a start (Amorós-Figueras et al. 2018).
Code lab
The lab builds a hybrid correction and explores tissue impedance.
| Part | What you do | Data |
|---|---|---|
| A | A distorted impedance space | synthetic |
| B | Affine and smooth corrections (Equation 1, Equation 2) | synthetic |
| C | Error and coverage (Equation 3) | synthetic |
| D | Fusing two estimates (Equation 4) | synthetic |
| E | Frequency-dependent impedance (Equation 5) | synthetic |
- Python notebook: labs/python/m18_lab.ipynb. Run it from the course repository after
pip install -r requirements.txt, or open it in Google Colab. - Data: none to download.
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. Two unbiased estimates with variances 1 and 1. Fused variance?
Answer
0.5 (Equation 4).
S2. Variances 1 and 4. Weight on the first?
Answer
0.8 (Equation 4).
S3. The gap to the nearest reference point doubles. Linear interpolation error bound?
Answer
Four times larger (Equation 3).
S4. How many unknowns in a 3D affine map?
Answer
Twelve: nine in \(A\), three in \(\mathbf b\) (Equation 1).
S5. A Debye relaxation with \(\tau\) = 1 µs. Frequency of strongest relaxation effect?
Answer
About \(1/(2\pi\tau) \approx\) 160 kHz (Equation 5).
S6. At very low frequency, what dominates complex conductivity?
Answer
The real conductivity \(\sigma\) (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.
- Dielectric imaging explained. Abeln and colleagues’ review, with two illustrative cases (Abeln et al. 2021). This module: dielectric properties.
- Anatomy against angiography. Rottner and colleagues’ comparison of pulmonary vein diameters with angiography (Rottner et al. 2021). This module: dielectric properties.
- Impedance in scar. Amorós-Figueras and colleagues’ impedance magnitude and phase in healthy and scarred myocardium (Amorós-Figueras et al. 2018). This module: dielectric properties.
Further reading
- Dielectric imaging. A review (Abeln et al. 2021); anatomical accuracy (Rottner et al. 2021).
- Tissue properties. A parametric model of dielectric spectra (Gabriel et al. 1996); impedance in infarct scar (Amorós-Figueras et al. 2018).
- Related modules. m16 (magnetic localization), m17 (impedance localization), m19 (motion and drift), m20 (geometry), m34 (software architecture).