Steering Inter-Cell Leakage to Set Hospital Weights in Over-the-Air Federated Clinical Models

Federated learning lets many clinics train one diagnostic model without pooling patient data. Each site sends model updates, and a server averages them. When the sites are wireless point-of-care devices, such as handheld ultrasound units in rural clinics or mobile screening vans on a shared cellular network, the averaging can happen in the radio channel itself. In this "over-the-air" aggregation, all devices transmit at once and the received signal is the sum of their updates. Existing designs assume a single isolated cell. In a real network, a device near a cell boundary is also heard by the neighbouring base station. The usual view treats that leakage as interference to be suppressed. Our observation is that when every cell trains the same model, the leakage is not noise. It is another copy of a useful update, added with a weight that the channel chose rather than the system designer.

That uncontrolled weight matters most in clinical-radiomic models, which combine features computed from medical images with a handful of clinical variables such as age or lab values. Image-derived features are known to encode the scanner and hospital that produced them, and in the TCGA glioma data deep networks can predict the acquisition site itself. A channel that quietly gives boundary clinics more or less influence is therefore changing which scanners the model fits. We propose to co-design device transmit power and how the server combines the signals received at different base stations. The goal is to make the effective per-site weights in the federated average match a weighting the clinician chooses, such as patient-proportional, equal per site, or worst-site-robust, while using leakage as extra received energy instead of nulling it.

Core Claim

In multi-cell over-the-air federated learning, the network's cross-cell channel matrix sets the weight each site receives in the global objective. Treating inter-cell interference as noise does not just add noise. It biases the model toward sites with favourable geometry. Joint power control and cross-cell combining can (a) remove that bias with a target-weight constraint and (b) do so at lower transmit power and fewer time slots than orthogonalising cells. Under site-heterogeneous radiomic data, this should show up as better worst-site AUC (area under the ROC curve) per unit of wall-clock time than both treat-as-noise aggregation and cell time-sharing.

Concretely, base station $c$ receives $y_c=\sum_k h_{ck} b_k g_k + n_c$. Here $g_k$ is device $k$'s model update, $b_k$ its transmit scaling, $h_{ck}$ the channel from device $k$ to cell $c$, and $n_c$ receiver noise of variance $\sigma^2$. The server forms $\hat g=\sum_c w_c y_c$, so device $k$'s effective weight is $\alpha_k=b_k\sum_c w_c h_{ck}$. We solve

$$ \min_{w,b}\ \sigma^2\|w\|^2 \quad \text{s.t.}\quad |\alpha_k-\pi_k|\le\epsilon,\ \ |b_k|^2 G_k\le P_k \ \ \forall k, $$

where $\pi_k$ is the target site weight, $\epsilon$ a tolerated weight error, $G_k$ a bound on update energy, and $P_k$ the device power budget. Treat-as-noise is the special case where $w$ selects one cell per device. Leakage then enters $\alpha_k$ as an uncontrolled term. The multi-antenna version replaces scalars with receive combining vectors, the cross-cell analogue of the receive beamformer in JDU-BF.

Positioning Against Prior Work

The closest work is task-aware wireless aggregation. Joint Downlink-Uplink Beamforming for Wireless Multi-Antenna Federated Learning optimises beamformers against a bound on the learning loss rather than signal-to-noise ratio, but it is single-cell and assumes zero inter-cell interference. One-Bit Over-the-Air Aggregation for Communication-Efficient Federated Edge Learning makes latency independent of device count using sign-based majority vote, also in one cell. We reuse their alternating-optimisation and truncated channel-inversion machinery. We differ in making the cross-cell weighting, not the noise, the primary design target.

Two works motivate the claim that weighting matters. Multi-cell load coupling (Resource Optimization With Load Coupling in Multi-Cell NOMA) shows that treating neighbour interference as static noise produces inefficient coupled equilibria. Site batch effects in TCGA (Advancing The Cancer Genome Atlas glioma MRI collections, Biased Data, Biased AI) show that site identity leaks into image features.

Clinical federated learning methods address non-identical site data but assume perfect digital links. These include FENDA-FL, mode-connectivity aggregation and differentially private clinical FL. Our contribution is to show the physical channel silently overrides their chosen weights. Interference graphs (vertex coloring for D2D) supply our strongest competitor, cell time-sharing.

On the clinical side, DL vs. conventional radiomics for IDH is a constraining result: handcrafted radiomics adds little over deep features. That result is why we place the deep imaging branch, the large one, on the analog channel.

Research Program

We will build a multi-cell simulator with hexagonal cell layouts in the style of the NOMA load-coupling testbed. Path loss and fading will be calibrated against DeepMIMO indoor channels as used in Neural Calibration. We first reproduce single-cell JDU-BF and OBDA on MNIST to validate the simulator.

Clinical tasks are partitioned by true acquisition site. The first is IDH genotype prediction on TCGA-GBM/LGG, using the CaPTk radiomic features and segmentations from Bakas et al. together with a small 3D convolutional imaging branch, age, and a fusion head. Sites are mapped to devices placed at controlled distances from cell edges.

The model is hybrid by design. The clinical branch and fusion head have few parameters, so they are sent digitally and error-free. The imaging branch is aggregated over the air. We then implement the weight-matching optimiser in the scalar and multi-antenna versions, and a sign-quantised variant compatible with OBDA. We will also analyse how the residual weight error $\epsilon$ propagates into the federated objective. This builds on the standard result that aggregation-weight mismatch converges to the optimum of a reweighted objective.

Baselines

The naive reference is per-cell treat-as-noise aggregation. Each base station inverts its own devices' channels as in OBDA's truncated inversion and forwards its sum to the server, ignoring leakage. It shows what the physics does by default.

The ideal reference is error-free FedAvg with the target weights $\pi_k$. This is the analogue of JDU-BF's "Ideal FL", which reached about 91% MNIST accuracy, roughly matched by JDU-BF in about 40 rounds (JDU-BF). On the clinical task, no federated or wireless result on TCGA glioma IDH was found on Lacuna. The nearest centralised figure is AUC 0.920 from combined DL and radiomic features in the IDH genotyping study. That study used an internal cohort plus TCGA, so we will re-train our centralised model on TCGA alone rather than borrow that number.

The strongest competitor is interference-graph time-sharing. We colour cells with Welsh-Powell so that adjacent cells aggregate in different slots, removing leakage exactly at a slot cost equal to the number of colours. A second competitor is single-cell JDU-BF applied independently per cell.

All methods share the same partitions, model, device power budgets, round count, and hyperparameter search budget. Each method's power and step-size knobs are tuned on a validation site set. Beating the strongest baseline means matching its worst-site AUC within confidence intervals while using fewer channel slots or less transmit energy, or exceeding it at equal slots.

The key ablation sets $\pi_k$ unconstrained, minimising noise only. This isolates whether gains come from weight control or merely from leakage-as-diversity. A second ablation sends the whole model, clinical branch included, over the air.

Why not just…?

Why not just time-share the cells so there is no inter-cell interference? This is the real threat. It recovers exact weights and costs only a constant factor in slots, typically three to seven colours. Our method must win on accuracy per slot or per joule, not on accuracy alone. If, at realistic power budgets, time-sharing's worst-site AUC per slot matches ours, we will report that the gain is not worth the cross-cell coordination. The sweep over cell density and device power is designed to find the regime, if any, where diversity from leakage pays.

Why not just send everything digitally on orthogonal channels? For hospitals on wired backhaul, that is the right answer, and we scope it out. For many wireless point-of-care devices, OBDA shows orthogonal latency grows linearly with device count while over-the-air aggregation stays constant. We do send the tiny clinical branch digitally, because uniform analog noise on it would be wasteful.

Why not just reweight sites in software after aggregation? The server receives only one sum per cell, so it cannot separate devices after the fact. Reweighting is possible only through the combining weights $w$ and transmit scalars $b$, which is exactly our optimisation.

Why not just treat interference as noise and rely on robustness of FedAvg? That is the naive baseline. The decisive check below tests whether its bias is clinically measurable.

Proposed Experiments

The first decisive check asks whether treat-as-noise aggregation measurably distorts the learned model. We will measure the total-variation distance between realised and target site weights. We will also compare held-out worst-site AUC against ideal FedAvg as boundary sites move closer to the cell edge. If the AUC gap is within seed noise, the clinical motivation fails and the work reduces to a power-saving result. We will say so.

The second check compares our optimiser against time-sharing and per-cell JDU-BF on worst-site AUC, mean AUC, and slots and energy to reach a target AUC. The sweep covers cell density, device power, and channel-estimate error. MNIST and CIFAR-10 with site-skewed label partitions provide a better-powered replication that is not limited by the 243-patient TCGA cohort.

The third check runs the ablations. Unconstrained-noise optimisation separates weight control from diversity. The all-analog model tests the hybrid split. A handcrafted-radiomics-only model tests whether, when the model is small, digital transmission simply wins, as Gap 3 in the research file suggests it might.