When Fading Picks the Patients: Over-the-Air Federated Learning for Cuffless Arterial Monitoring

Wearable pulse sensors can now estimate blood pressure and arterial stiffness continuously. Examples include optical sensors that measure blood-volume pulses at the fingertip or toe, and piezoelectric patches over the wrist artery. Models that generalize across people need data from many wearers. That data is sensitive: the pulse waveform works like a fingerprint. A separate body of wireless research shows that when many devices transmit analog model updates at the same time on the same frequency, the radio channel sums them for free. The receiver's own thermal noise can then serve as the random perturbation that differential privacy requires, so each device does not need to add noise itself. This technique is called over-the-air computation, and the resulting privacy guarantee is called channel-induced differential privacy. Nobody has tried it on physiological signals. Every published evaluation uses MNIST, FEMNIST, CIFAR, Iris, or synthetic regression.

This note argues that the interesting problem at this intersection is not the thermal noise. It is who gets dropped. In practical over-the-air schemes, a device whose radio link is in a deep fade cannot reach the power needed to align with the others, so it is excluded from that round. On a body-worn sensor, deep fades are not random. They depend on arm posture, movement, and distance from the receiver, and these in turn correlate with the wearer's current hemodynamic state and with who the wearer is (for example, sedentary older hypertensive patients versus active young users). The channel may therefore quietly under-sample exactly the physiology the model most needs to learn. We propose to measure this bias and to fix it with a participation rule that balances clinical groups. The rule is paired with a compact physiological feature parameterization, which keeps the model small enough that channel noise gives a meaningful privacy budget.

Core Claim

In over-the-air federated training of cuffless blood-pressure and stiffness regressors, fading-driven client exclusion that is correlated with posture and physiology produces per-group error gaps. At privacy budgets where thermal noise alone costs little accuracy, these gaps are larger than the accuracy cost of the noise. We will test two fixes together. The first bounds dropped weight separately for each coarse clinical group, not only in aggregate. The second trains a regressor on low-dimensional pulse features rather than raw waveforms. Together they should close most of the gap while keeping the energy and bandwidth advantages reported for channel-native privacy.

The claim is falsifiable in a specific way. If realistic posture–fading correlation produces no measurable group gap, the balancing rule is unnecessary. The contribution then shrinks to a first characterization of channel-induced privacy on physiological regression, and we would report exactly that.

Grounding in Prior Work

Several papers motivate the privacy need. Generalizable Blood Pressure Estimation from Multi-Wavelength PPG Using Curriculum-Adversarial Learning trains a blood-pressure model on four-wavelength fingertip photoplethysmography (PPG, optical blood-volume pulse). It needed a gradient-reversal layer because a discriminator could readily identify subjects from the features. This is evidence that pulse waveforms are biometric, and that anonymizing them is weak protection. The same paper also shows that segment-level train/test splits inflate accuracy, and that subject-level splits give much larger errors. We therefore treat each subject as a separate federated client and evaluate only on held-out subjects.

On the sensing side, Thin, soft, wearable system for continuous wireless monitoring of artery blood pressure builds a wrist patch that measures local pulse wave velocity over 15 mm. It feeds the measured features to a per-device gradient-boosted model and runs about 48 hours on a coin cell. This battery budget makes transmit energy a first-class metric for us. The patch also needed cold-pressor and hand-grip stress tests to produce enough blood-pressure range for training, which argues against purely on-device learning. Photoplethysmography Signals and their Correlation with Peripheral Artery Disease identifies a handful of toe-PPG shape features that track the ankle-brachial index, such as width at half amplitude (R² = 0.295) and maximum rising slope (R² = 0.292). It also discards 26% of recordings for poor quality. Together these results suggest that a small set of hand-defined features carries much of the clinical signal, and that noise tolerance is limited. Cuffless Blood Pressure Estimation Algorithms for Continuous Health-Care Monitoring likewise meets the medical-device accuracy standard (AAMI) by fusing pulse arrival time with shape features. We reuse this feature-first design, because a feature-based regressor has far fewer parameters than a CNN on wavelet images.

On the wireless side, Certified Characterization of Privacy, Participation, and Convergence in Over-the-Air Federated Learning supplies our core mechanism, a method called CARES. It shows that a single receive scaling factor $\eta$ couples three things at once: noise, privacy sensitivity, and which clients get truncated. It derives certified envelopes on exclusion probability under Rayleigh fading and imposes an aggregate dropped-weight constraint. It then selects $\eta$ adaptively under zero-concentrated differential privacy accounting (zCDP), gaining 8.9 points on FEMNIST. Its dropped-weight constraint is blind to which clients drop, and that is the gap we fill.

Communication and Energy Efficient Wireless Federated Learning With Intrinsic Privacy shows that random sparsification lowers sensitivity and bandwidth. It reports 600 versus 2,000 subcarriers, and 2.55×10¹¹ J versus 3.70×10¹¹ J for an artificial-noise baseline on CIFAR-10. Our feature parameterization is a domain-specific alternative to that generic compression, and we test the two against each other directly.

Two further papers bound what we can assume. Rethinking Federated Learning Over the Air: The Blessing of Scaling Up shows that mutual-information leakage falls as $O(1/N)$ with the number of clients, and that power control becomes unnecessary near $N=100$. Our cohorts are 87–2,362 subjects, which sits in that regime, so blind transmission is a serious baseline. That paper also assumes phase synchronization, which wearables may not maintain. Privacy for Free: Wireless Federated Learning via Uncoded Transmission With Adaptive Power Control and AirMixML establish SNR regimes where privacy costs nothing, and warn that high dimensionality makes channel noise destructive. Both results push us toward small models.

Research Program

The build has three parts. The first is a client model. Each subject's device extracts per-beat features: pulse transit or arrival times, rising slope, half-amplitude width, fall-time fractions, and, for multi-wavelength PPG, the same features per band. It then trains a small regressor (a two-layer network of a few thousand parameters) to predict systolic and diastolic pressure, or the ankle-brachial index. Updates are clipped to a fixed norm for sensitivity and sent as uncoded analog symbols with truncated channel inversion, following CARES.

The second part is a channel simulator in which fading is not independent of physiology. Each device's large-scale gain depends on a latent posture/activity state. That state also shifts the device's current blood-pressure distribution, and a parameter $\rho$ sets the correlation strength. We sweep $\rho$ from zero, the standard assumption, to strong. We add Doppler-driven channel estimation error and a phase-misalignment parameter, because the file flags synchronization as untested on wearables. No body-area channel measurements appear in the research file, so this is a controlled simulation, and we say so plainly. Measured on-body channel traces would be the natural follow-up.

The third part is the group-balanced participation rule. Clients carry a coarse group label $g$, such as a pressure range bin derived from their own cuff calibrations. The server needs per-group participation totals without learning individual labels. Clients therefore transmit one-hot group indicators over the air in the same slot, so group counts arrive as a noisy sum that is covered by the same privacy accounting. The rule replaces CARES's aggregate constraint with a per-group constraint:

$$\max_{g}\; \frac{\sum_{i \in g} w_i \,\mathbb{1}[\,|h_i|^2 < \eta\, c_i\,]}{\sum_{i \in g} w_i} \le \beta ,$$

Here $w_i$ is client $i$'s aggregation weight, $h_i$ its fading coefficient, $c_i$ the channel-inversion threshold that depends on its power limit, and $\beta$ the tolerated dropped fraction per group. The bandit that selects $\eta$ keeps CARES's reward but must satisfy this tighter constraint.

A second, optional lever reweights surviving clients by inverse estimated inclusion probability within each group. This raises effective sensitivity, and we account for that cost explicitly.

Baselines

The naive reference points are a centralized, non-private model trained on the same features with subject-level splits, and a per-device local-only model. The centralized model bounds achievable accuracy. The local-only model tests whether collaboration helps at all. For the published centralized number on the multi-wavelength cohort, the curriculum-adversarial model in Generalizable Blood Pressure Estimation reports a systolic MAE of 14.2 mmHg and a diastolic MAE of 6.4 mmHg under strict subject-level splitting. We will reproduce it as the centralized upper reference, and our non-private federated model should come within about 1 mmHg of it. ACNN-BiLSTM reports 1.67 and 1.15 mmHg on a closely related cohort. Its splitting protocol is not stated in the file, and the tenfold gap to the subject-split result suggests leakage, so we treat it as non-comparable unless re-run with subject-level splits.

The strongest privacy competitor is digital federated learning with on-device Gaussian noise and standard DP-SGD accounting. This mirrors the artificial-noise baseline (WFL-PDP) in Communication and Energy Efficient Wireless Federated Learning With Intrinsic Privacy. It uses the same model, the same clipping norm, the same target $(\epsilon,\delta)$, and the same total transmit-energy and subcarrier budget. It is the strongest competitor because it has no fading-driven exclusion at all, so it is immune to the bias we target. It only pays in energy and bandwidth. No reported result was found on Lacuna for digital DP federated learning on PPG blood-pressure data.

The strongest over-the-air competitors are three. The first is CARES itself with its aggregate dropped-weight constraint. The second is fixed-$\eta$ truncated inversion, over which CARES reports +8.9 points on FEMNIST and +2.7 on MNIST. The third is blind, CSI-free transmission as in Rethinking Federated Learning Over the Air. For compression, we run the randk sparsification of PFELS on a raw-waveform CNN, with the compression ratio tuned as in that paper.

Every method receives the same hyperparameter search budget, the same rounds, and the same simulated channel draws. Two ablations isolate our ingredients. The first swaps the per-group constraint for the aggregate one while keeping everything else fixed. The second swaps feature inputs for raw waveforms with randk compression.

Beating the baselines means the following at matched $\epsilon$ and matched energy. Worst-group MAE should be lower than CARES's, with the difference exceeding seed variance. Overall MAE should be no worse than CARES's. The method should come within a stated margin of digital DP federated learning while using fewer subcarriers and less energy.

Why not just…?

Why not just use digital federated learning with on-device noise? This is the real threat. The evidence against it comes from vision simulations, namely the energy and subcarrier savings in PFELS, and from the observation in Free Privacy Protection for Wireless Federated Learning that uncorrected bit errors plus artificial noise hurt convergence in noisy links. Neither result has been shown for small physiological models, where payloads are tiny and energy differences may be negligible. Our matched-budget comparison settles it. If digital DP federated learning matches accuracy at similar energy for a few-thousand-parameter regressor, we will say that over-the-air privacy is not worth it at this model size, and report the model size at which it becomes worth it.

Why not just keep CARES's aggregate dropped-weight constraint? If fading were independent of physiology, the aggregate constraint would suffice, and our $\rho=0$ condition should show that. The case for per-group constraints rests entirely on the $\rho>0$ experiments. We report the $\rho$ at which the two diverge, and whether that $\rho$ is physically plausible.

Why not just use generic compression such as randk instead of physiological features? Random sparsification is proven for over-the-air sensitivity reduction in PFELS. However, dicrotic-notch and slope information is subtle, and the toe-PPG study shows that a few features already capture ABI correlations. Our ablation compares the two at equal transmitted dimension. If randk on raw waveforms wins, we adopt it.

Why not just transmit blind with many clients, since power control becomes unnecessary? Rethinking Federated Learning Over the Air shows this works near 100 clients with synchronized phases. We run it directly, and also under our phase-misalignment sweep, which is the condition the file flags as untested on wearables.

Proposed Experiments

The main testbed is the four-wavelength fingertip PPG cohort (180 subjects, 66 hypertensive, 200 Hz) used by Generalizable Blood Pressure Estimation, with one client per subject and five-fold subject-level cross-validation. Groups are normotensive versus hypertensive and posture state. The second testbed is the UCSD toe-PPG cohort (2,362 patients) from Photoplethysmography Signals and their Correlation with Peripheral Artery Disease. It gives a large-$N$ regime for ABI regression, with PAD severity as the group variable. If access permits, the 87-subject radial patch data serves as a third, small-$N$ setting where truncation should matter most.

We report subject-level systolic and diastolic MAE, worst-group MAE, the gap between groups, zCDP $\epsilon$ converted to $(\epsilon,\delta)$, total transmit energy, and subcarriers. Each configuration is run over ten channel seeds.

The decisive checks are three. First, whether a group gap appears under CARES as $\rho$ rises, which tests the core claim. Second, whether the per-group rule closes that gap without raising overall MAE or requiring a larger $\epsilon$, which tests the fix. Third, whether any over-the-air variant beats matched-budget digital DP federated learning on energy by a margin that matters for a 48-hour coin-cell device. If the first check comes back null at plausible $\rho$, the paper becomes a negative-but-useful result: channel-induced privacy transfers cleanly to physiological regression, and group balancing is unnecessary.