Listening for the Hinge: Side-Channel Analysis of Lithotripsy Shockwaves to Detect Through-Cracks in Calcified Arteries

Intravascular lithotripsy (IVL) prepares heavily calcified arteries for a stent by firing small shockwaves from emitters inside a low-pressure balloon. The waves pass harmlessly through soft tissue and crack the stiff calcium. Computer models show that the treatment pays off mainly when one specific outcome occurs: a crack that runs all the way through the calcium ring and splits it, so the ring opens like a hinge. Shallow cracks of the same length help much less. Clinicians currently cannot tell, while they are firing pulses, which of these outcomes they have produced.

We propose to treat each lithotripsy pulse the way a security researcher treats a computer that is unintentionally leaking information through its power draw or sound. Every pulse already produces byproduct signals that nobody reads: the acoustic echo, the pressure transient in the balloon, and the electrical discharge waveform. Reading these signals pulse by pulse, and comparing each one against the first pulse in the same artery, should reveal the moment the calcium ring loses its hoop stiffness. The result would be a crack-state monitor for IVL, built with the methods of acoustic side-channel attacks, aimed at an open problem in lesion preparation.

Core Claim

Our claim is that splitting the calcium ring is a large, abrupt mechanical change, and that it should be readable from signals the IVL system already produces. A closed rigid ring resists outward load through hoop stress. A ring split by a through-crack has lost that load path and can only bend open, which is a much softer and lower-frequency mode of response. The in silico study of Impact of lesion preparation-induced calcified plaque defects motivates the target. It builds 22 parametric finite element models calibrated on optical coherence tomography (OCT) images from 262 Disrupt CAD-III patients. In those models, through-cracks give relative lumen gain of up to 51% for large plaques, against 16–25% for non-penetrating cracks of equal length. The same study finds that crack length matters more than width and that crack location does not change lumen gain.

That last result is useful for sensing. If location does not matter, the monitor only needs to answer one question, "has the ring split?", and does not need to locate the crack. The question is binary, carries most of the clinical value, and corresponds to a step change in the ring's stiffness. That combination is what makes a side-channel reading plausible.

The methodological lens comes from SoK: Acoustic Side Channels. It frames leakage as an inverse problem, a map $S: I \times M \to O$ from hidden inputs $I$ and conditions $M$ to observations $O$. It judges reconstruction by stability (small changes in the signal cause small changes in the answer) and uniqueness (different hidden states give different signals). The SoK reports that machine-learning decoders recover 3D-printer instructions from motor sounds with 78–100% accuracy and keystrokes with up to 94%. It does not examine medical actuators. Here the hidden input is the crack state, and the "attacker" is the clinician. We borrow one specific device from side-channel practice: differential analysis, meaning comparing traces from the same device rather than reading absolute signals. Writing $O_n$ for the multichannel trace of pulse $n$ in a given artery, the monitor scores the differential trace

$$\Delta_n = O_n - O_1,$$

which cancels the patient-specific transmission path through tissue, balloon, and catheter. It keeps only what has changed in the lesion since the first pulse. Our hypothesis is that through-crack formation produces a discontinuity in $\Delta_n$ across pulses that is larger than the pulse-to-pulse noise. Intact rings and non-penetrating cracks should instead produce a gradual drift.

Research Program

The work has three parts, and each is a precondition for the next.

The first part is an identifiability analysis by simulation. We will extend the parametric geometries of the in silico study, using 90° and 270° calcium arcs, intact rings, non-penetrating defects of varying length, and through-cracks, from static pre-cut meshes to explicit dynamic simulations. In these simulations a pressure pulse is applied inside the fluid-filled balloon at 4 atm. The original study names exactly this simplification as a limitation: it treats lithotripsy as static geometric pre-cuts and does not model wave propagation. We will record three simulated observables: the reflected pressure at the balloon wall, the balloon pressure transient, and a far-field acoustic signal.

From these simulations we will compute how separable crack states are as a function of sensor noise. We will use classifier-estimated mutual information between crack state and $\Delta_n$, and the side-channel "success rate versus number of traces" curve. The analysis also answers a uniqueness question. Can a long non-penetrating crack mimic a short through-crack? If it can, the monitor is ill-posed, and we will report that as a negative result rather than proceed.

The second part is a bench and ex vivo study, which settles whether the simulated signal survives real noise. We will build calcified ring phantoms with calcium-loaded mineral rings set in soft polymer tubing. Their crack states will be controlled by pre-scoring, and the rings will then be treated with a commercial IVL catheter. We will also treat human calcified arterial segments ex vivo. After every pulse group, ground truth will come from OCT and micro-CT. Calcium will be segmented with the volumetric classifier of Automated Volumetric Intravascular Plaque Classification Using OCT, which reports 81.15% overall accuracy on ex vivo cryo-image validation. Every fracture call will be confirmed by an expert reader. The study records three channels: a hydrophone on the catheter shaft or a skin-coupled contact sensor, the inflation-line pressure transducer, and a current probe on the emitter discharge.

The third part is a decision study. A stopping rule built on the monitor ends therapy once the through-crack probability passes a calibrated threshold. We will check whether this rule matches the through-crack rate and stent expansion of a fixed full-dose protocol while using fewer pulses. Stent expansion will be measured on phantoms and ex vivo segments by OCT after deployment at 14 atm, matching the in silico protocol.

The security direction contributes method, not a threat model. A privacy reading is possible: echoes "leak" plaque architecture to an eavesdropper. That attacker is implausible in a catheterization lab, and we do not pursue it. We likewise set aside the enclave-leakage questions raised by TEE-X and SNPeek. They are real, but they do not touch the IVL open problem. One lesson from SNPeek does carry over. Its finding that implementations designed to be constant-time still leak is a reminder that a system built to be "quiet" toward soft tissue can still be informative about the hard tissue it disrupts.

Baselines

The naive reference is the dose itself. The baseline predicts through-crack state from cumulative pulse count and calcium arc alone, with no sensing, fit as a logistic model on the same bench and ex vivo specimens. This is what clinicians effectively do today. The monitor is worthless unless it beats pulse count by a clear margin in area under the receiver operating characteristic curve (AUROC) under leave-one-specimen-out evaluation.

The strongest competitor is direct imaging. An OCT pullback after each pulse group, segmented by the classifier of Automated Volumetric Intravascular Plaque Classification Using OCT, is the gold standard for seeing a crack. That paper reports a median 100% plaque-type accuracy in leave-one-pullback-out in vivo validation and 81.15% on ex vivo cryo-images. It does not report crack detection, and we found no reported result on Lacuna for automated through-crack detection from OCT during IVL. We will therefore measure it ourselves on the same specimens, scored blind against micro-CT. The comparison is fair because both methods see the same lesions and are evaluated against the same ground truth. The acoustic monitor "beats" imaging if it reaches non-inferior through-crack AUROC, within a pre-registered margin of 0.05, with no catheter exchange or flush. We do not expect it to be more accurate than imaging.

A third baseline is a balloon compliance probe. It reads only the inflation-pressure response to each pulse, or to a small volume step, and is the cheapest physical proxy for hoop stiffness. Two ablations isolate the key ingredients. The first scores absolute traces $O_n$ instead of differential traces $\Delta_n$, which tests the claim that within-artery comparison is what makes the problem stable across specimens. The second replaces the physics-informed features with a generic spectrogram convolutional network trained with the same data and tuning budget. Every model gets the same hyperparameter search budget and the same specimen-level splits.

Why not just…?

Why not just image with OCT after every few pulses? Because imaging interrupts therapy. It needs a catheter exchange, a contrast flush, and time, and the file reports that even automated OCT analysis takes workflow time measured in tens of minutes to an hour. That is too slow for a pulse-by-pulse decision. Imaging is still the real competitor on accuracy, which is why it is our strongest baseline rather than a straw man. If the acoustic monitor cannot come within the pre-registered margin of OCT, the honest conclusion is that it can serve at most as a trigger for when to image.

Why not just test balloon compliance? This is the most serious threat. A split ring is softer, and the inflation line already carries a pressure sensor. We do not argue around it. The pressure channel is part of the monitor and is also run alone as a baseline. If pressure alone matches the full multichannel monitor, the finding is still useful and much cheaper: the shockwave itself serves as the stiffness probe. The acoustic channels then earn their place only by a measured AUROC gain. One reason to expect such a gain is that at 4 atm the balloon may only weakly load the ring, so the quasi-static compliance change could be small compared with the change in high-frequency reflections. The bench data will decide.

Why not just fire the full pulse dose every time? Fixed dosing ignores the outcome that matters. In the in silico results, lumen gain depends on whether a through-crack forms, not on how many pulses were fired. The parallel-wire literature points the same way: Effectiveness of the parallel wire technique cites 100% versus 41.2% 24-month patency when a single controlled crack is created. A fixed dose also gives no warning when the ring never splits, which is exactly the lesion in which the stent will under-expand. The pulse-count baseline measures how much information dosing alone carries.

Why not just train a black-box classifier on raw recordings? With tens of ex vivo specimens, a generic model is likely to learn specimen identity instead of crack state. The SoK shows that device fingerprinting from acoustics reaches 98% accuracy, so specimen identity is an easy shortcut. The differential-trace and physics-feature ablations test this directly. If the black-box model generalizes equally well across held-out specimens, we will adopt it.

Proposed Experiments

The decisive check comes first and runs in simulation. If the explicit dynamic models show that through-cracks and long non-penetrating cracks produce overlapping $\Delta_n$ distributions at realistic noise levels, the program stops there. That would answer a uniqueness question the inverse-problem framing poses but the original in silico study never asked. If the states separate, the bench phase tests transfer from simulation to phantoms. The ex vivo phase then reports through-crack AUROC, the number of pulses until detection, and calibration, all under leave-one-specimen-out evaluation and stratified by 90° and 270° arcs. Stratifying by arc matters because the in silico global-benefit metric is higher for extensive calcification ($\mu_{270}=192$ versus $\mu_{90}=144$), so that is where a monitor matters most.

The final experiment compares the monitor-guided stopping rule with full-dose and pulse-count rules on through-crack rate, post-stent lumen gain measured by OCT, and pulses delivered. Success means equal expansion with fewer pulses, or earlier identification of lesions that will not split. Either outcome would turn the pulse's byproduct signals into the clinician's first view of whether the calcium ring has split.