Rechargeable medical implants, such as the pulse generators that drive deep brain stimulation, are charged through the skin by an external coil. How well that works depends on how well the two coils line up. When the implant tilts, rotates, or migrates in its pocket, less of the magnetic field passes through it, and the charger has to push more current to deliver the same energy. More of that energy becomes heat in the tissue under the pad, and charging sessions get longer. The patient is caught in the same trade-off building and car-cabin controllers face: faster charging means more energy delivered, while cooler, shorter sessions mean more comfort. Here, though, how steep the trade-off is depends on coil alignment, which changes over months. We propose a charging controller that estimates alignment from signals the charger already measures. The user or clinician sets an allowed rate of too-hot sessions, and the controller spends that "discomfort budget" to maximize delivered energy. It recalibrates the budget as the implant drifts, and it is evaluated with long, within-patient designs rather than short bench tests or one-time surveys.
The core claim is that controllers which bound comfort violations with statistical guarantees learned from a short calibration window will quietly break their guarantees when the underlying physical coupling drifts. Adding an alignment estimate plus recalibration that weights recent data more heavily restores the promised violation rate at little cost in delivered energy. A second claim is methodological. Cross-patient, single-timepoint comparisons of charging burden mix up alignment effects with implant depth and body build. Only within-patient panels can say how much of the burden misalignment actually causes.
What the literature gives and what it leaves open
Three literatures motivate the problem. The first establishes that alignment drives efficiency. Xiao Lu et al. survey how inductive links lose power quickly as distance and misalignment grow. Hongyang Li, Daoqing Su et al. report that flipping or rotating an implanted pulse generator in its pocket causes complete charging failure. Xiaoran Fan, Longfei Shangguan et al. show that pulse, breathing, and muscle motion keep disturbing transcutaneous links. All three characterize the physics or specific failures. None treats alignment as a slowly drifting state that a controller must track across months while also respecting a comfort constraint.
The second literature supplies the control method. Jicheng Shi, C. Salzmann, and Colin N. Jones build a data-driven predictive controller (DAD-DPC) for buildings. It uses split conformal prediction, a distribution-free way of setting error margins from held-out data, so that the fraction of time outside the comfort band stays near a user-chosen level α, and it trades that allowance for energy savings. We reuse this "spend a violation budget" formulation, with local skin heating as the violation and delivered charge as the objective. Their guarantee assumes calibration data resemble future data. Their longest real deployment ran 40 days, including a 19-day learning period. That window is too short to test whether the guarantee survives the slow drift that implant pocket migration produces. Wenjie Xu et al. similarly keep cumulative discomfort below a threshold with contextual Bayesian optimization. Our contribution is to make coupling the context and to make the calibration drift-aware.
The third literature constrains how we evaluate. The limited-follow-up studies show short windows mistaking early gains for durable ones. Sohal et al. rely on six-month surrogate markers. Yang et al. predict arrhythmia recurrence from 3–6 month tracking, which misses late recurrence. Yunhan Zheng et al. report an EV-charger spending effect that shrinks between their two windows. The cross-sectional studies make the complementary point. Duncan B. Clark et al. explicitly flag that their mediation chain needs longitudinal confirmation. Xianglong Liu et al. show that panel fixed effects with an instrument can recover effects that pooled snapshots miss. We borrow the within-unit fixed-effects logic, with the patient as the unit, and replace the instrument with small randomized perturbations of charger power, which the device permits.
Research Program
The controller estimates the coupling coefficient $k$, a number between 0 and 1 measuring how much of the transmitter's field links the receiver, once per session. It uses the transmitter-side reflected impedance. Misalignment changes this impedance because the reflected term scales with the square of the mutual inductance, and no extra implant hardware is needed. It then chooses transmitter current $I$ to solve
$$\max_{I} \; \mathbb{E}[P_{\text{del}}(I, \hat k)] \quad \text{s.t.} \quad \Pr\big[T_{\text{skin}}(I, \hat k) > T_{\max}\big] \le \alpha,$$
where $P_{\text{del}}$ is power delivered to the battery, $T_{\text{skin}}$ is peak skin temperature under the pad, $T_{\max}$ is the comfort limit, and α is the allowed fraction of too-hot sessions. The temperature margin comes from conformal residuals of a heating model conditioned on $\hat k$. Those residuals are recomputed with exponentially decaying weights over past sessions, so calibration tracks pocket drift instead of averaging over it. When $\hat k$ falls below a floor, as in a flipped device, the controller stops raising current and gives a repositioning cue in the prescriptive style of CARE.
We would build this in three stages. First, a long-horizon stress test on BOPTEST checks the drift-aware calibration in a well-understood domain. Second, a tissue-phantom bench with a motorized receiver reproduces tilt, lateral offset, flipping, and slow scripted migration, with thermocouples under the pad. Third, a clinical study uses existing charger logs from patients with rechargeable implants. Those logs record session duration, transmitter current, and completion. The retrospective log analysis is followed by a prospective within-patient study using micro-randomized power levels inside safe limits.
Baselines
The naive reference point is the usual clinical behavior: fixed transmitter power with a hard thermal cutoff, run on the same phantom scripts and the same patients. To beat it, we must deliver more energy per session minute at equal or lower over-threshold rate. No reported result was found on Lacuna for this baseline.
The strongest existing method is DAD-DPC with static conformal calibration (Shi et al.). It reports energy savings at a 5% violation target of 30.1%, 11.2%, and 27.1% on three BOPTEST cases and 20.5% on the Polydome, where it reached a 2.5% violation rate after 19 of 40 days. It is the strongest competitor because it already provides the exact guarantee we care about. On BOPTEST we run it unchanged, with its published settings as the reference and the same tuning budget for our variant. On the implant tasks we give it the same heating model and data and remove only the coupling context and the drift weighting. Beating it means matching its savings or delivered energy while keeping the realized violation rate within α at every evaluation horizon, in the regime where its realized rate drifts above α.
The hardware baseline uses coupling-tolerant transmitters: the omnidirectional LCC design of S. Liu et al., the 3D coil arrangement of Hao Qiang et al., and the PT-symmetric circuit of Zhenghao Zhu et al., which keeps efficiency stable across coupling changes. We build the PT-symmetric variant on the phantom bench and run it with the naive controller. No reported result was found on Lacuna for any of these at implant scale.
The advice baseline follows Azaria et al.. CARE's prescriptive advice cut cabin energy by 17% while information alone was not significant. We compare a reposition cue alone, without current control, and an information-only display.
Two ablations isolate the key ingredients: coupling-aware control with static calibration, and drift-aware calibration without the coupling context.
Why not just…?
Why not just use a coupling-tolerant circuit so the controller is unnecessary? This is the most serious threat. If a PT-symmetric or omnidirectional transmitter keeps efficiency flat across the tilts implants actually experience, alignment stops driving heating and our controller adds little. The file reports that such circuits stabilize efficiency across varying coupling. It gives no implant-scale numbers and no evidence that they handle a flipped device, where Li et al. observe total failure. The phantom bench settles this directly. If the PT-symmetric transmitter under naive control matches our controller's delivered energy and violation rate across the full migration script, we would drop the control claim and keep only the evaluation contribution. The two approaches are also complementary, so the bench includes the combination.
Why not just tell patients to reposition the coil? CARE shows that prescriptive advice works where information does not (Azaria et al.), so a cue-only arm is a fair competitor. Repositioning cannot undo pocket rotation beneath the skin or rhythmic micro-motion (Fan et al.), and it asks effort of the patient every session. The cue-only arm tells us how much of the gain comes from advice versus current control.
Why not just run a longer trial with the existing controller? Longer follow-up is necessary but not sufficient. A longer trial would reveal that static calibration drifts but would not fix it. A longer cross-sectional or pooled comparison would still confound alignment with implant depth, which is the problem Clark et al. flag for their own design. The within-patient randomized design is what turns long follow-up into a causal estimate.
Proposed Experiments
The first decisive check runs on BOPTEST. We run static-calibration DAD-DPC and the drift-aware variant over horizons well beyond 40 days, injecting slow parameter drift into the building models to mimic envelope or equipment aging. We track the realized violation rate against α at successive horizons and energy relative to the rule-based default. The prediction is that static calibration exceeds α once drift accumulates, while the drift-aware version stays within it at an energy cost of a few points. If static calibration holds, the drift premise is wrong and we say so.
The second check runs on the phantom bench. We cross every controller with scripted tilt, offset, flipping, and months-equivalent migration. We measure delivered energy per minute, transmitter energy per delivered watt-hour, and the over-threshold session rate against α. We also check whether $\hat k$ flags a flipped receiver before any heating occurs.
The third check is the clinical panel. From charger logs we estimate the effect of coupling on session length and early termination in two ways: pooled across patients at one timepoint, and with patient fixed effects over time. Where the prospective arm exists, we add micro-randomized power levels. A large gap between the two estimates would show that snapshot comparisons mislead in this setting. The primary outcome is whether the controller keeps the over-threshold rate within α at 1, 3, 6, and 12 months. Secondary outcomes are patient-reported heating discomfort and charging adherence, on the 7-point and binary scales used by AutoTherm.
References
- Wireless Charging Technologies: Fundamentals, Standards, and Network Applications
- Recharging Difficulty With Pulse Generator After Deep Brain Stimulation: A Case Series of Five Patients
- Towards flexible wireless charging for medical implants using distributed antenna system
- Disturbance-Adaptive Data-Driven Predictive Control: Trading Comfort Violation for Savings in Building Climate Control
- Data-driven adaptive building thermal controller tuning with constraints: A primal-dual contextual Bayesian optimization approach
- A prospective evaluation of cardiovascular magnetic resonance measures of dyssynchrony in the prediction of response to cardiac resynchronization therapy
- A predictive model using left atrial function and B‐type natriuretic peptide level in predicting the recurrence of early persistent atrial fibrillation after radiofrequency ablation
- Effects of electric vehicle charging stations on the economic vitality of local businesses
- Adolescent Executive Dysfunction in Daily Life: Relationships to Risks, Brain Structure and Substance Use
- Causal association between PM2.5 and metabolic syndrome in the Chinese elderly population-insights from a cohort study of CHARLS
- OMNIDIRECTIONAL WIRELESS POWER TRANSFER EXPERIMENT BASED ON LCC RESONANCE NETWORK
- Three-dimensional transmitter for omnidirectional wireless power transmission: design and experiment
- Parity–Time Symmetric Model and Analysis for Stable Multi-Load Wireless Power Transfer
- Advice Provision for Energy Saving in Automobile Climate Control Systems
- AutoTherm: A Dataset and Ablation Study for Thermal Comfort Prediction in Vehicles