Hospitals and water-monitoring labs test bacteria against antibiotics with the disk diffusion assay. A paper disk loaded with drug sits on a lawn of bacteria spread on agar, the drug diffuses outward, and the diameter of the clear ring where bacteria could not grow tells you how susceptible the strain is. Labs would like to train a shared model that predicts resistance across their combined isolates, but they cannot pool patient-linked data. That points to federated learning: each lab trains on its own data and only model updates leave the building. The labs also report at very different speeds, since plates incubate overnight and results arrive in irregular weekly batches. That points to asynchronous federated learning, where the server folds in updates as they arrive rather than waiting for everyone.
The obvious recipe is to take an existing method that already handles both slow clients and non-identical client data and run it on susceptibility data. Our claim is that this recipe misreads the problem. In disk diffusion, labs differ in two ways that a generic method cannot tell apart. Some labs see different bacteria (biological heterogeneity). Others pour agar slightly deeper, inoculate slightly denser, or read ring edges differently, which shifts every ring diameter in a systematic way (measurement drift). Generic drift correction treats both as "client updates disagree with the consensus." Measurement drift is a persistent, lab-specific bias, not an optimization artifact, and fighting it with gradient corrections wastes the correction and contaminates the shared model. We propose to absorb measurement drift into a small per-lab physical calibration model derived from diffusion physics. The parameters of that model are pinned down by the quality-control reference strains every accredited lab already runs. Drift correction and staleness handling then apply only to the shared biological model.
Core Claim
Ring formation follows from Fick's law. A point dose of drug spreading in a thin agar layer produces a concentration that falls off with distance. The ring edge sits where that concentration drops to the minimum inhibitory concentration (MIC, the lowest drug concentration that stops growth) at a critical time when the lawn becomes too dense to clear. For a disk of drug mass $m_0$ in agar of depth $h$, drug diffusion coefficient $D$, and critical time $T_0$, the point-source solution gives
$$ r^2 = 4 D T_0 \,\ln\!\frac{m_0}{4\pi D T_0 h} \;-\; 4 D T_0 \,\ln \mathrm{MIC}, $$
where $r$ is the ring radius. Writing $a_{l,j}$ for the intercept of drug $j$ in lab $l$ and $b_l = 4 D T_0$ for the slope gives the per-lab observation model $r^2_{l,j} = a_{l,j} - b_l \ln \mathrm{MIC}_j + \varepsilon$. Agar depth enters only through the intercept. Inoculum density and incubation conditions enter through $T_0$, so they change both slope and intercept. The relation is affine in squared radius against log MIC, not in raw diameter. That gives a specific, testable prediction about which coordinate system a calibration layer should use.
The shared model $f_\theta$ predicts an isolate's vector of log MICs from organism identity, specimen source, and the zones it has for other drugs, which also lets it impute untested drugs. Each lab keeps a private calibration $(a_{l,\cdot}, b_l)$ mapping predicted MICs to its own observed rings. Two kinds of data identify the calibration. CLSI and EUCAST quality-control strains have published reference MIC ranges, and labs routinely re-run them, as the inter-laboratory reproducibility work by Qingquan Shi et al. and Baodi Shang et al. emphasizes. The QC results anchor each lab's calibration. Calibration parameters never leave the lab and are never averaged, so they cannot go stale. Only $\theta$ goes through asynchronous aggregation with drift correction.
Relation to Prior Work
The aggregation and correction machinery is reused, not invented. FedBuff supplies the server design. It collects $K$ asynchronous updates in a buffer, down-weights stale ones, and keeps compatibility with secure aggregation. FedAC supplies three pieces: cosine-similarity weighting of updates, prospective momentum at the server, and a correction applied at every local step to keep client models near the global trajectory. FedAC is the closest existing work. It treats all client disagreement as drift to be corrected and was evaluated only on vision and language benchmarks. Our proposal differs by splitting off a structured, physically interpretable component of the disagreement before correction is applied.
That split also bears on FedAC's similarity weighting. A lab with a genuine calibration offset produces updates misaligned with the server trajectory even when its biology is perfectly representative. Down-weighting such a lab throws away good data.
SCAFFOLD motivates the correction idea through per-client control variates that cancel gradient dissimilarity. Its stateful per-client memory is a known cost. We do not need it here because the persistent part of each lab's difference lives in the explicit calibration. Server-side proxy-data methods such as FedCLG align updates using a reference dataset held at the server. QC strains play an analogous anchoring role, but they are measured inside each lab under that lab's own conditions. That is exactly what makes them informative about measurement drift rather than only about gradient direction.
On the application side, the surveillance studies of Helen M. Sanchez et al., Ewa Karwowska et al., and the decade-long keratitis series of Epidemiological and Antimicrobial Resistance Trends in Bacterial Keratitis show that disk diffusion data are siloed and heterogeneous in organism mix. Machine Learning for Antimicrobial Resistance shows the value of prediction but assumes centralized data. Both motivate the setting. None separates measurement from biology across sites.
Research Program
We will build three things.
The first is a plate simulator that generates ring diameters from the Fickian relation. It adds lab-specific depth, inoculum, incubation, and reader noise. It also includes deliberate violations: hydrophobic or slowly diffusing agents whose rings are too small for their true potency, and hazy bacteriostatic edges. The simulator lets us dial biological heterogeneity, using a Dirichlet split over organisms and resistance prevalence, independently of measurement drift. The key experiments require that independence.
The second is the training algorithm. It is FedBuff-style buffered aggregation with FedAC's per-step correction, both applied only to $\theta$. Each lab jointly fits its calibration $(a_l, b_l)$ during local steps, with a penalty tying the calibration to its QC-strain estimate.
The third is a small wet-lab ring trial in which participating labs run a shared panel of reference strains and drugs under deliberately varied agar depth and inoculum density. This tests whether the squared-radius affine form actually holds on real plates.
Charge carrier recombination enters as a stress test, not as a pillar. Photocatalytic antibacterial materials such as the thermolysis ZnO of Enhanced Photocatalytic and Antibacterial Performance of ZnO Nanoparticles kill bacteria through reactive oxygen species. These form only when light-generated electrons and holes separate before they recombine. That paper reports an MIC of 30–50 µg/mL against E. coli, measured by broth dilution rather than diffusion. In well-diffusion screens of such materials, the lab's illumination intensity becomes a further measurement variable. It changes apparent potency through carrier separation, not through transport, and nanoparticles diffuse poorly enough that the Fickian form may break down. We will simulate a federation of materials-screening labs with lab-specific illumination. The aim is to see whether the calibration layer degrades gracefully when its physics is wrong. We also test whether adding a single illumination-scaling parameter per lab recovers performance. This tells us how far the approach extends beyond standard antibiotics.
Baselines
The naive reference is local-only training, where each lab fits its own model. Next to it is centralized pooled training on all simulated data with known lab identity, which serves as an upper bound that ignores privacy.
The asynchronous baseline without explicit correction is FedBuff with polynomial staleness down-weighting. The strongest existing method is FedAC. On CIFAR-10 with Dirichlet $\alpha = 0.1$, FedAC reports 70.6% top-1 accuracy against 63.1% for FedBuff. We will first reproduce this pair on that benchmark to confirm our implementations are faithful before trusting any comparison on new data. For the susceptibility task itself, no reported result was found on Lacuna for any federated method, so all numbers will be produced in this study. All methods will use the same simulated federations, the same client arrival traces, the same model architecture, and an equal hyperparameter search budget per method.
Two ablations isolate the key ingredient. In the first, the per-lab calibration is learned freely with no QC anchoring. In the second, the calibration is anchored but affine in raw diameter rather than squared radius.
Beating FedAC means two things. Our method must achieve lower very-major error (calling a resistant isolate susceptible, the clinically dangerous mistake) at matched wall-clock time. It must also achieve higher essential agreement (predicted MIC within one doubling dilution of truth). Both gains must appear across at least three measurement-drift levels, and the advantage must shrink toward zero as measurement drift goes to zero. If the advantage does not shrink, our explanation for it is wrong.
Why not just…?
Why not just subtract each lab's QC-strain offset and then run FedBuff or FedAC? This is the most serious threat. QC correction is standard practice, and it might capture most of the benefit. The physics predicts it should fall short only when drift changes the slope $b_l$ (inoculum and incubation through $T_0$) rather than just the intercept (depth), and when the offset is applied in diameter rather than squared radius. The simulator lets us vary depth-only versus inoculum drift. If offset subtraction matches our method under inoculum drift too, we will report that the contribution reduces to "use QC anchors, in the right coordinates." That would be a smaller claim than this proposal makes.
Why not just use personalized federated learning with a free per-lab output head? A free head cannot distinguish a lab whose isolates are genuinely more resistant from a lab whose rings run small. It will absorb real biology into calibration and erase the surveillance signal the federation exists to find. The unanchored-calibration ablation tests this directly. We expect it to look fine on held-out isolates from the same lab but to misestimate cross-lab resistance prevalence. We will measure prevalence error against simulator ground truth.
Why not just run synchronous training, since a consortium has few labs? With few labs, FedBuff's buffer can approach the total number of clients and asynchrony buys little, so this alternative is plausible. We will replay realistic batch-arrival traces with overnight incubation and weekly reporting and compare wall-clock time to target agreement against synchronous FedAvg with the same calibration layer. If synchronous training wins at consortium scale, the calibration idea still stands on its own and the asynchronous framing is dropped.
Decisive Checks
The central experiment varies measurement drift while holding biological heterogeneity fixed, and the reverse. The prediction is that our advantage over FedAC grows with measurement drift and vanishes without it. A second check withholds QC data from a subset of labs. Performance should degrade gracefully for those labs only, not for the federation. A third check tests whether FedAC's similarity weights, run without calibration, systematically down-weight labs with large calibration offsets. That would show directly that generic drift correction is discarding representative data.
The ring trial settles whether squared radius is linear in log MIC across depths and inocula on real plates. If it is not, the parametric layer is replaced by a monotone learned calibration and the claim narrows accordingly. The photocatalytic stress test reports how much of the gain survives when ring size reflects light-driven carrier separation rather than diffusion.
References
- Epidemiological and Antimicrobial Resistance Trends in Bacterial Keratitis: A Hospital-Based 10-Year Study (2014–2024)
- A Novel Transposon Tn7709 Harbors Multidrug Resistance Genes in a Pathogenic Aeromonas media Strain QST31
- Federated Learning with Buffered Asynchronous Aggregation
- Efficient Asynchronous Federated Learning with Prospective Momentum Aggregation and Fine-Grained Correction
- SCAFFOLD: Stochastic Controlled Averaging for Federated Learning
- Accelerating Hybrid Federated Learning Convergence Under Partial Participation
- Antibiotic Resistance of Escherichia coli Isolated from Conventional, No Antibiotics, and Humane Family Owned Retail Broiler Chicken Meat
- Microbiological Hazards Associated with the Use of Oligocene Waters: A Study of Water Intakes in Warsaw, Poland
- Machine Learning for Antimicrobial Resistance
- Enhanced Photocatalytic and Antibacterial Performance of ZnO Nanoparticles Prepared by an Efficient Thermolysis Method