Fiber-Aligned Spatial Reuse for Galvanic Implant Networks

Implanted and body-worn devices can talk to each other by pushing tiny currents through tissue. This method, called galvanic coupling, uses the body itself as the wire. When many such devices share the same frequencies, each link's "interference zone" decides which links can safely transmit at the same time. Current spectrum-sharing methods draw that zone as a circle around each transmitter. Tissue does not behave that way. In white matter and heart muscle, current travels much farther along fibers than across them. The EEG forward-modeling literature reports along-fiber conductivity up to nine times the across-fiber value (Review on solving the forward problem in EEG source analysis).

We propose to build the map of who interferes with whom from the tissue's direction-dependent conductivity. That map would come from diffusion MRI, a scan that measures the direction water moves in tissue, or from cardiac fiber atlases. The payoff has two sides. A circular map either wastes reuse opportunities across fibers or causes collisions along them. The question is how much shared bandwidth a dense implant network gains, and how many collisions it avoids, when its scheduler uses the true elongated interference zones.

Core Claim

The claim is narrow and testable. Implant networks will pack more and more electrodes into neural recording and closed-loop stimulation systems. At that density, the spectrum they share is underused mainly because the interference model is wrong, not because devices are scarce.

The simplest case makes the geometry concrete. Take a uniform tissue whose conductivity differs by direction, described by a 3×3 conductivity matrix $\Sigma$. A point current source of strength $I$ produces a potential at displacement $r$ of

$$u(r) = \frac{I}{4\pi\sqrt{\det\Sigma}\,\sqrt{r^{\top}\Sigma^{-1}r}}.$$

This is the standard solution of $\nabla\cdot(\Sigma\nabla u) = -I\delta$ after rescaling coordinates. The consequence is that the set of positions where interference exceeds a threshold is an ellipsoid, not a sphere. Its axes scale with the square root of the principal conductivities. At the 9:1 ratio reported for white matter, the zone reaches about three times farther along fibers than across them.

A scheduler that uses a circle therefore faces a choice. It can use a radius large enough to be safe along fibers and forbid many harmless cross-fiber pairings. Or it can use a smaller radius and let along-fiber links collide.

Real tissue is not uniform. Skull, cerebrospinal fluid and scar tissue change the picture. So the full method replaces the closed form with coupling computed from a finite-element model: a simulation that splits the tissue into small elements, each with its own conductivity matrix. The closed form is kept as an explicit, cheaper competitor.

Research Program

Physics layer. We build patient- or atlas-specific volume conductor models with tissue-specific conductivity matrices. The pipeline is the one Butson et al. used for deep brain stimulation, which maps diffusion tensors onto a finite-element mesh (Predicting the Effects of Deep Brain Stimulation with Diffusion Tensor Based Electric Field Models). That work motivates ours. It showed that an isotropic model predicted a side-effect threshold of −1 V, while the diffusion-tensor model predicted −4 V, which matched the patient. Direction-dependent conduction therefore changes how far electrical effects reach by clinically meaningful amounts.

We reuse that pipeline for communication rather than therapy. That shift is the gap noted in the research file: tissue-modeling papers study diagnosis and stimulation, and body-area-network papers such as Spectrum Sensing in a Cognitive Body Area Network: Detection of a Bonded Channel in the MICS Band do not model tissue conductivity at all.

Cost layer. Finite-element solves are slow. The electrocardiography calibration study needed about 200 minutes for 100 forward evaluations (Using Bayesian Optimization to Calculate Conductivity in an Electrolytic Fluid for the Forward Problem in Electrocardiography). That is far outside the millisecond-to-second control loops used by spectrum managers such as Iris and xDiff.

We avoid putting the solver in the loop, using a method taken from EEG modeling. Reciprocity, reviewed by Hallez et al., gives the coupling from every possible transmitter location to a given receiver in one forward solve. We therefore run one solve per receiver site offline, accelerated with algebraic multigrid, a fast solver for large sparse systems. At run time we only look up precomputed coupling values. When electrodes shift, we update coupling by interpolation rather than remeshing, following Butson et al.'s dynamic interpolation.

Calibration layer. The absolute conductivity of living tissue is uncertain. The research file lists in vivo parameter determination as unresolved. Inverse problems of this type are also formally non-unique in three dimensions (On nonuniqueness for Calderón's inverse problem). We do not try to recover the full conductivity field.

Instead we treat the physics-predicted coupling matrix as the prior mean of a Gaussian process, a statistical model that is updated as measurements arrive. We then correct it with a small number of pilot transmissions between node pairs. The Gaussian-process choice follows the Bayesian optimization used for conductivity by Rojas-Velázquez et al. and for Wi-Fi spatial reuse by INSPIRE. Only a handful of bulk parameters are fit, such as the along/across ratio and the skull scale factor. The link budget depends on these, so the non-uniqueness result does not block this step. The electrocardiography study's mismatch between simulated potentials (−2.57 to 3.37) and measured ones (−1.6 to 1.66) is a warning about exactly this scale error. It is why calibration is not optional.

Access layer. The calibrated coupling matrix defines a weighted interference graph. We then use existing distributed access rules unchanged. The main one is the spatial congestion game of Distributed Spectrum Access with Spatial Reuse, where selfish users choosing channels reach a stable outcome. We also use an underlay rule with power control and multipacket reception from Interference-based Optimal Power-efficient Access Scheme for Cognitive Radio Networks. Keeping the access algorithms fixed isolates the contribution of the interference model.

Baselines

The naive reference is static orthogonal scheduling. Every implant gets its own time or frequency slot, as in the TDMA baselines of Multi-Agent Deep Stochastic Policy Gradient for Event Based Dynamic Spectrum Access. In that work, TDMA reached under 60% event-reporting rate under correlated traffic, while the learned policy reached about 92%. Static scheduling has zero interference by construction, so it bounds reliability from above and reuse from below. Beating it means higher total throughput at equal or lower collision rate once there are more than a few nodes.

The strongest existing method is Chen and Huang's distributed spatial congestion game with a circular interference range. That paper reports equilibrium utility within 8% of the centralized optimum. It has only been evaluated in a 250 m × 250 m radio setting; no reported result found on Lacuna for any spectrum access method in a tissue volume conductor. We run this baseline on the same simulated tissue, node placements and traffic as our method. Its interference radius is tuned by grid search over the same held-out placements used for tuning ours, and it receives the same pilot-measurement budget to set that radius. Beating it means a Pareto improvement: more concurrent links or throughput at matched collision probability for primary links.

The key ablation replaces the finite-element coupling with the closed-form ellipsoid above. The fiber direction comes from diffusion-tensor data, with no heterogeneity. A second ablation drops the physics prior and fits the Gaussian process to pilot measurements alone. This mirrors INSPIRE's model-free local optimization.

Why not just…?

Why not just tune the radius of an isotropic interference model? A single radius cannot match an ellipse with a 3:1 axis ratio. Any radius is too conservative in one direction or too permissive in the other. The geometry fixes the trade-off, and tuning cannot remove it. The Butson threshold result shows the isotropic error is large in tissue. Our tuned-radius baseline measures how much this costs in actual reuse. If the tuned circle comes within a few percent of our method, the anisotropy argument fails for this application, and we will report that.

Why not just measure every pairwise coupling directly? This is the most serious threat. With $N$ nodes, full probing costs on the order of $N^2$ pilot transmissions. It must be repeated whenever electrodes migrate or the patient's posture changes. It also spends implant energy, which is the binding constraint in the energy-limited regime studied by Anwar et al. The measurement-only ablation settles the question. We plot reuse efficiency against the number of pilots. The physics prior earns its place only if it reaches a given accuracy with substantially fewer pilots, or holds up after simulated electrode movement where the measurement-only model goes stale.

Why not just use the closed-form ellipsoid and skip finite elements? This may be enough in deep white matter, where local fiber direction dominates. It is likely to fail near the skull and cerebrospinal fluid. Hallez et al. report that the skull is 10–80 times less conductive than soft tissue and redirects current sideways. The ablation is stratified by anatomical region. If the closed form matches finite elements everywhere, we adopt it, since it is simpler.

Proposed Experiments

Simulation benchmark. The first experiment is a simulation benchmark built on diffusion-tensor head models, following Butson et al.'s pipeline, and on cardiac models with fiber architecture. In each, we place 16 to 128 galvanic transceivers at recording and stimulation sites. Traffic follows the bursty event-driven model of Kassab et al. We measure edge-level accuracy of the interference graph, separately for false negatives (missed interference) and false positives (forbidden but harmless pairs). We also measure spatial reuse factor, aggregate throughput, collision rate on protected links, and pilot energy.

Movement stress test. The second experiment moves electrodes by millimeters and updates coupling by interpolation. It compares staleness under the physics prior with staleness under the measurement-only model.

Physical check. The third experiment tests in a bench phantom with measurable fiber alignment, such as ex vivo muscle, using transmit–receive electrode pairs. First we confirm that measured coupling contours are elongated with the predicted axis ratio. Then we confirm that the calibrated prior reproduces them after the few-parameter Bayesian fit. The EDGAR torso-tank data, from the electrocardiography calibration study, is used only to check our scale calibration against a known mismatch. It is an isotropic electrolyte and cannot test the anisotropy claim.

Decisive check. The decisive result is the pilot-budget curve. The physics-prior graph must beat both the tuned circular congestion game and the measurement-only Gaussian process at low pilot budgets and after movement. If it does not, the conclusion is that tissue anisotropy is a modeling curiosity rather than a spectrum resource.