Repeated On-Body Cold Pulses as a Wearable Vascular Stress Test

Many disease-causing defects stay hidden while the body is at rest. Backup pathways and compensating mechanisms keep resting measurements looking normal, and the defect only shows when the system is pushed. Cell biologists exploit this on purpose. They stress cells with a drug, radiation, or simulated ischemia, and the hidden defect appears. Wearable sensors do the opposite. They record passively and treat any disturbance as noise to throw away. This note proposes giving a small wearable patch its own stressor. The patch would deliver short, mild, precisely timed cooling pulses to the skin, and because the device knows exactly when each pulse happened, it can average the blood-vessel response across many pulses. Disturbances that are not tied to the pulses, such as motion, average out, while the vessel response survives.

The two-sentence version is this. A wearable that applies its own small, repeatable cold challenges can measure how a person's blood vessels react, and it can separate that reaction from motion noise by averaging over the known pulse times. We expect this to give more reliable per-person measurements than either resting recordings or naturally occurring stress events. The averaging is cheap enough that the computing budget is negligible, so the hard engineering problem is the energy cost of the cooling element, not the classifier.

Core Claim

The clearest existing evidence that a controlled physical stressor exposes something resting signals do not is the forearm cold-challenge study Differential Sensitivity of IPG and PPG Sensors to Temperature-Induced Peripheral Vasoconstriction. It compared two sensors:

  • Photoplethysmography (PPG) is the optical pulse sensor in smartwatches. Real ice dropped PPG pulse amplitude by 41% (p < 0.001), while sham ice raised it by 27%.
  • Impedance plethysmography (IPG) measures pulsing blood volume electrically and reaches deeper arteries. Its amplitude did not change significantly.

The authors' reading is that cold constricts the superficial capillary bed but not the deeper arteries. That study plays a motivating role here, and it also defines the protocol we must match. Its limits are equally clear. It used hand-placed ice and a benchtop Biopac MP36R rack. It ran a single challenge on 21 healthy people. It did no on-device analysis.

Our claim is that this kind of challenge can move onto the body as a repeatable, automated protocol, and that repetition is what makes it valuable. The patch would deliver N mild cooling pulses at known onset times $t_1,\dots,t_N$. Let $x(t)$ be the ratio of PPG to IPG amplitude. Using the ratio cancels changes that affect both vascular beds, such as heart rate or blood pressure. The per-person response estimate is the stimulus-locked average

$$\hat r(\tau) = \frac{1}{N}\sum_{i=1}^{N} x(t_i+\tau) - \bar x_{\text{pre}},$$

where $\tau$ is time since pulse onset and $\bar x_{\text{pre}}$ is the pre-pulse baseline. If motion artifacts and spontaneous fluctuations are not tied to the pulse schedule, their contribution to $\hat r$ shrinks roughly as $1/\sqrt{N}$, as in any average of independent noise. Randomizing the inter-pulse intervals makes this independence assumption much safer.

This addresses an open tension in the literature. Wearable ECG work treats activity-induced change as noise to reject; Deep Learning-Based Signal Quality Assessment in Wearable ECG Monitoring builds a deep network for exactly that purpose. Yet Beyond Hearing found that noisy free-living data was more useful for pretraining than clean clinical data. A known stimulus clock offers a principled way to tell physiological response from artifact without a large gatekeeper network. The response is whatever is time-locked to the stimulus.

We are explicit about scope. The cell-level evidence motivates the idea but does not transfer directly. A BAG3 variant that looks normal until proteasome inhibition Investigation of a dilated cardiomyopathy–associated variant in BAG3 is one example. A TTN carrier whose ion-channel defects surface only under ischemia A pathogenic titin missense mutation is another. Neither says anything about forearm vessels. This project tests the engineering and statistical premise in a tractable vascular setting, and it does not screen genetic carriers.

Research Program

The device. We will build a forearm patch with three parts:

  • a small thermoelectric (Peltier) element with a skin thermistor;
  • a PPG front-end beside it and an IPG electrode pair over the radial artery, matching the sensor geometry of the benchtop study;
  • an accelerometer, used only for evaluation and not as an input to the main estimator.

The stimulus. Pulses are mild, on the order of a few degrees of skin cooling for tens of seconds. That is far gentler than ice, and it is a deliberate safety and energy choice. Inter-pulse intervals will be randomized.

The energy problem. Thermoelectric cooling draws on the order of watts while active. That is orders of magnitude above the roughly 382 µW average system power that ECG-TCN needs for continuous inference on a coin cell. The actuator therefore runs into the battery-volume limit described in Lightweight holographic near-eye display system, where batteries exceed half of device weight.

We will treat energy per usable response estimate as a primary metric. We will also trade pulse depth against the number of pulses, since shallower pulses cost less energy but need more repetitions to reach the same precision.

The analysis. Analysis runs on a microcontroller. The stimulus-locked average and a few derived features are cheap: response depth, time to minimum, and recovery time constant. The recovery curve matters because the prior hardware work classifies static windows and has no notion of a recovery trajectory.

Baselines

Naive reference: resting features. This is the do-nothing point. It uses resting PPG and IPG features from the same patch over the same session length, with no stimulation: pulse amplitude, amplitude variability, and pulse arrival time. It tells us whether stimulation adds anything at all.

Passive natural-stress baseline. This is the strongest cheap competitor. Over a matched multi-day free-living period, we would detect spontaneous events such as ambient temperature changes and posture shifts. We would then compute the same response features locked to those events, with event counts matched to the number of active pulses. Beating it means higher test-retest reliability and higher known-group discrimination at equal recording time.

Learned resting-signal baseline. This represents the claim that a large model needs no stimulus. We will train a self-supervised encoder in the style of PiMT and Bidirectional-Mamba from Beyond Hearing on our free-living patch data. It will be fine-tuned under the same labels, cross-validation splits, and hyperparameter search budget as our feature model. That paper reports 87.6% average F1 on DailySense but about 58% F1 on unseen users. The unseen-user figure is the relevant one, because our evaluation is entirely across people. No reported result found on Lacuna for any learned model on vascular cold-reactivity.

Strongest existing protocol: the benchtop ice challenge. We will rerun the protocol of Differential Sensitivity of IPG and PPG Sensors on the same participants on separate days: ice versus sham on a Biopac rack. The reported reference is a 41% PPG amplitude drop with no significant IPG change. The patch "beats" the benchtop protocol if its averaged response correlates strongly with the benchtop response within person, and if its day-to-day reliability is higher than that of the single benchtop challenge.

Ablations. Two ablations isolate the key ingredients:

  • Hidden timing. Run the same patch and the same pulses, but have the estimator search for responses without being given pulse times. This isolates the contribution of the known stimulus clock.
  • PPG only. Drop the IPG ratio and use PPG alone. This isolates the value of the deep-bed reference channel.

Why not just…?

Why not just run the benchtop or clinic challenge once? This is a real threat. If one supervised ice test is as reliable across days as many patch pulses, the case for the wearable shrinks to accessibility alone. The benchtop study reports a single session per person and gives no day-to-day reliability, so the comparison is open. Our repeated-measures design settles it directly by comparing intraclass correlation (ICC) across days for the benchtop protocol and for the patch.

Why not just use naturally occurring stress? The free-living hardware already exists and costs no actuator energy. Its weakness is that natural events vary in strength and timing, and they are tangled up with motion. We test it head-to-head with matched event counts. If its reliability matches ours, we will report that the actuator is unnecessary.

Why not just learn from resting signals with a big model? The cellular evidence shows compensation erasing resting differences, for example BAG3 myofibrils that are indistinguishable at baseline. That evidence is only an analogy for vessels. The learned resting baseline is the direct test. Its reported collapse on unseen users makes it a weak cross-person competitor, though larger free-living pretraining could narrow that gap.

Why not stream raw data to a phone and analyze it there? For this task, compute is not the bottleneck. A time-locked average costs almost nothing compared with the actuator. Streaming would still be needed only to adapt pulse timing in real time. We will report radio energy against actuator energy so that readers can see which one dominates.

Proposed Experiments

Cohort and data. The study will have three arms:

  • a healthy cohort, tested on three separate days for reliability;
  • a known-group arm of people with clinically diagnosed cold-induced vasospasm, compared with age-matched controls;
  • a provoked-artifact arm, in which participants walk and move their arms during pulses.

The file lists no public dataset with controlled on-body perturbation and simultaneous PPG and IPG, so data collection is part of the contribution. BUT QDB and the Shelab wearable ECG database are ECG resources. They only inform how we label motion-corrupted segments.

Decisive checks. Five results would decide the question:

  • The averaged response should replicate the benchtop direction of effect: PPG drops and IPG stays stable.
  • Cross-day ICC of patch features should exceed both the passive baseline and the single benchtop challenge.
  • Known-group discrimination, measured as AUC with participant-level cross-validation, should exceed the resting and learned baselines.
  • In the walking arm, the hidden-timing ablation should degrade sharply while time-locked averaging stays close to its seated estimate.
  • Estimation error should fall roughly as $1/\sqrt{N}$ with pulse count; a flatter curve would mean artifacts are tied to the stimulus and the independence assumption fails.

Every result will be reported alongside energy per reliable estimate.