Wastewater-based epidemiology — inferring community disease burden from viral RNA measured in sewage — is sold as an unbiased escape hatch from passive clinical surveillance, which only sees patients who feel sick enough to seek care. But the sewer itself is not a neutral pipe. Two largely ignored filters sit between a community's true infections and the numbers an analyst sees. The first is physical: viruses travel attached to micron-scale solids in a moving fluid, and fluid dynamics research shows that in pipes, small particles get trapped against the wall by near-wall turbulence and lift forces, with trapping rates that depend on flow speed and geometry. The second is statistical: labs gate which samples get expensive genomic sequencing using arbitrary rules of thumb, such as the cycle threshold cutoff Ct < 32, which can silently discard exactly the low-concentration samples in which an emerging variant first appears. Neither filter is corrected anywhere in current practice.
This proposal argues that both filters are correctable, and that fixing them requires the union of three literatures that have never touched. Near-wall mesh refinement from computational fluid dynamics (CFD) supplies the mechanism for the physical filter: resolving the thin, steep-gradient layer of fluid hugging the pipe wall — where deposited material actually exchanges with the flow — is what makes particle trapping and re-release computable. Passive surveillance bias research supplies the diagnosis: the epidemiology literature has carefully documented how selective observation warps inferred incidence, from Zika's twelve-to-fifteen-month silent circulation before first clinical detection to chlamydia's apparent decline that masked a real rise. ROC-based threshold optimization supplies the cure for the statistical filter: instead of an arbitrary Ct gate, choose the sequencing threshold that maximizes detection of true community signals at a defensible false-alarm cost.
Core Claim
The claim is specific and falsifiable. Current wastewater surveillance treats the sewer network as a well-mixed black box and the lab gate as a fixed constant. We claim that (a) near-wall-resolved multiphase CFD can turn the sewer reach into a calibrated instrument, yielding a transport kernel — a function mapping what the community sheds today to what the treatment plant measures later and weaker — whose shape is predicted by hydraulic covariates (velocity, pipe diameter, bend curvature); and that (b) replacing the fixed Ct gate with a cost-weighted ROC-optimized threshold recovers early low-frequency variants that the fixed gate drops. Both sub-claims are directly testable against existing public benchmarks, and neither has been attempted in the literature we surveyed on Lacuna.
The point is not that wastewater surveillance is broken. Sun et al. showed sewage RNA tracks clinical positivity with Spearman's ρ = 0.97 (p < 0.001) in Yantai City and leads clinical detection of several Omicron lineages by roughly one week — an impressive baseline (Beyond clinical data: the role of sewage monitoring in tracking pandemic trends of SARS-CoV-2). The point is that this performance is achieved despite the two unmodeled filters, and the same paper explicitly names the missing physics as its main limitation: the absence of information on sewage physicochemistry and environmental factors affecting viral stability and detection. We propose to build exactly that missing piece.
Research Program
The program has three components, each converting an existing method from a neighboring field into this setting.
First, the transport kernel. Qin and Zhou demonstrated, in a 5-meter horizontal pipe on a 12-million-cell hexahedral mesh with boundary layer refinement, that for micron particles wall deposition is governed by near-wall turbulent diffusion — smaller particles deposit more at higher gas velocities, the "1 μm anomaly" — and that lift forces are essential for particles below roughly 15 μm (Numerical simulation of pneumatic conveying characteristics of micron particles in horizontal pipe). Their companion elbow study found that centrifugal forces drive particles to the outer bend wall, with pressure loss minimized near a bend diameter ratio R/d ≈ 4 (Numerical simulation of flow characteristics in vertical elbows of pneumatic conveying). Both studies are dry gas–solid pneumatic conveying; nobody has run this machinery for dilute liquid–solid sewer flow carrying pathogen-laden flocs. We will: build near-wall-resolved (y+ < 1, in the dimensionless wall-distance metric that indexes how deeply the first grid cell penetrates the viscous sublayer) multiphase simulations of representative sewer reaches across sewer-realistic velocities (roughly 0.3–2 m/s, not the 1–15 m/s of pneumatic lines), floc size classes, and bend ratios, and extract from each a kernel K(τ) of transit delays and a wall-exchange term. Formally, the plant measurement is modeled as
$$c_{\text{plant}}(t) = \int K_\theta(\tau)\, c_{\text{shed}}(t-\tau)\, d\tau, \qquad \int K_\theta(\tau)\,d\tau < 1,$$
where θ collects the hydraulic covariates of the reach and the sub-unity integral is the wall-trapping attenuation the black-box models currently ignore.
Second, fusion into the statistical layer. Ouyang, Law, and Doss built dlmwwbe, a dynamic linear model with Kalman filtering that smooths noisy twice-weekly wastewater series and estimates the lag between sewage concentrations and healthcare-worker infections — but they fit that lag empirically per site and note it varies substantially across sewer sheds (dlmwwbe: Dynamic Linear Model for Wastewater-Based Epidemiology). We will replace the free per-site lag parameter with the CFD-derived kernel, predicted from measurable hydraulics rather than fitted, so it transfers to sheds never seen in training. This is the same move the ascertainment-bias literature made for clinical data: Cameron et al. showed that decoupling observation effort (screening volume) from true disease overturned a naive reading — apparent stability in female chlamydia notifications after 2008 was screening saturation, while true incidence rose from 1,800 to 3,200 per 100,000 between 2001 and 2011 (Estimating incidence of sexually transmissible infections in Australia). We are decoupling observation physics from true signal.
Third, the gate. Sun et al. sequenced only wastewater samples with Ct < 32, applied as an unexamined rule (Beyond clinical data). The threshold-optimization literature says arbitrary cutoffs on continuous signals discard statistical power and can be selected against you: Liu et al.'s retinal imaging study is the cleanest demonstration — a global density measure gave near-chance AUC of 0.564 while a locally optimized measure gave 0.816, because averaging and fixed thresholds hide subclinical signal (Quadrant asymmetry alteration of retinal degeneration in thyroid-associated ophthalmopathy). And the standard Youden index, $J = \text{Sensitivity} + \text{Specificity} - 1$, weights misses and false alarms equally, which is wrong here: missing an emerging variant costs far more than sequencing an extra sample. We will optimize the cost-weighted criterion
$$J_c = C_{\text{FN}}\cdot \text{Sensitivity} + C_{\text{FP}}\cdot \text{Specificity},$$
where $C_{\text{FN}}$ and $C_{\text{FP}}$ are the costs of a missed emergence and a wasted sequencing run, swept explicitly so decision-makers can see the whole curve rather than one point. This follows the spirit of medical cutoff work like Luo et al.'s multicenter osteosarcoma study, which used Youden maximization to turn continuous inflammatory markers into risk stratification with validated survival separation (A new model of preoperative systemic inflammatory markers predicting overall survival of osteosarcoma: a multicenter retrospective study), but adapts the objective to asymmetric public-health costs.
Baselines
The do-nothing reference is the published Yantai pipeline of Sun et al.: raw RT-qPCR concentrations from treatment-plant samples, flow-adjusted but with no in-sewer transport model, and the fixed Ct < 32 sequencing gate (Beyond clinical data). Its reported numbers are the bar: Spearman's ρ = 0.97 (p < 0.001) against clinical positivity, lineage diversity exceeding clinical sampling (t = 6, df = 4, p < 0.05), and roughly one week of lead time on lineages DY.2, DY.4, and BA.5.2.48. We beat this baseline if, on held-out data, we detect designated low-frequency lineages earlier than one week before clinical confirmation, or detect lineages the fixed-gate pipeline misses entirely, without degrading the sewage–clinical correlation.
The strongest existing competitor is dlmwwbe itself: the same dynamic-linear-model machinery we use, but with per-shed lags fitted empirically rather than predicted from hydraulics (dlmwwbe). It is the strongest because it already solves the smoothing-and-lag problem statistically, so our contribution must show that physics beats statistics, not that modeling beats nothing. No reported result found on Lacuna for dlmwwbe's predictive accuracy on the Minnesota datasets, so we will re-run it ourselves on those same datasets — the wastewater series (March 2022–February 2023, twice-weekly ORFlab/Nlab targets) and wastewaterhealthworker (flow-adjusted sewage paired with confirmed healthcare-worker cases) — with an equal tuning budget and held-out sheds. Beating it means lower out-of-sample error on held-out healthcare-worker case counts and, critically, lag estimates that transfer to sheds excluded from fitting, which empirical per-site constants cannot do by construction.
The key ablation swaps the near-wall-resolved transport kernel for a well-mixed plug-flow delay: same pipeline, same optimizer, same data, but no boundary-layer trapping term. If the near-wall physics is the active ingredient, this ablation should recover most of the delay but little of the attenuation and its hydraulic dependence; if the ablation does just as well, the CFD is decoration and we will say so.
Why not just…?
Why not just let the Kalman smoother absorb the transport effects? Because a fitted lag is a constant, not a function of hydraulics. Ouyang et al. themselves report that optimal lags vary significantly across sewer sheds, which is exactly the symptom of unmodeled physics: each site gets its own patch constant. The Zika study shows what this costs in inference — Massad et al. had to back-calculate, from the ratio of observed to true infections, that Zika circulated silently for 12 to 15 months before passive detection, and that ~8,600 reported French Polynesian cases actually represented about 43,000 infections (On the origin and timing of Zika virus introduction in Brazil). A smoother fitted on biased observations inherits the bias; only a mechanistic kernel predicts beyond the fitted regime. Our held-out-shed experiment is designed to settle this directly, and if the empirical-lag DLM transfers as well as our kernel, the simpler method wins and the proposal's physical half is falsified.
Why not just lower the Ct gate (or sequence everything)? Lowering the gate uniformly trades a known cost for unknown gain: it multiplies sequencing volume across all samples to reach the few that matter, and sequencing budget is the binding constraint in real programs. The Liu et al. result is the caution — a poorly chosen threshold on a continuous signal can sit at near-chance discrimination (AUC 0.564) while a properly chosen one on the same underlying data reaches 0.816 (Quadrant asymmetry alteration of retinal degeneration in thyroid-associated ophthalmopathy). We will run the brute-force alternative as a comparison arm: a gate sweep at fixed budget will show whether any single fixed Ct value matches the cost-weighted optimum's detection of low-frequency clades. If one does, that is a useful practical finding and the ROC machinery collapses to "use Ct < X" — we will report it as such.
Why not just correct with a simple flow normalization or positivity ratio? The chlamydia record is a controlled experiment in how naive corrections fail. Cameron et al. explicitly evaluated the naive reading of Australian notification data, which suggested incidence was falling; the conditional-probability model that decoupled testing effort from infection showed true incidence rising and male incidence nearly doubling from under 4% to almost 8% (Estimating incidence of sexually transmissible infections in Australia). A flow-dilution correction in our setting commits the same error structurally: it corrects for volume but is blind to wall trapping (the "1 μm anomaly" of Qin and Zhou shows attenuation can increase with velocity, opposite to any dilution intuition) and to the gate-induced selection bias. This alternative is cheap, so we will implement it as a third comparison arm, and expect it to fail specifically on reaches where CFD predicts strong near-wall deposition.
Proposed Experiments
Experiment one is the CFD campaign: near-wall-resolved (y+ < 1, verified against a coarser wall-function variant as a mesh-convergence control) dilute liquid–solid simulations of straight and bent sewer reaches, following the 12-million-cell refinement methodology of Qin and Zhou but at sewer velocities and with viral-floc size classes (Numerical simulation of pneumatic conveying characteristics...), sweeping bend ratio around the R/d ≈ 4 optimum identified in the elbow study (Numerical simulation of flow characteristics in vertical elbows...). Deliverable: kernels K(τ) and attenuation coefficients as functions of θ, with the prediction that empirical lag heterogeneity across sheds correlates with these hydraulic covariates — the first decisive check.
Experiment two is the signal-correction test on the Minnesota datasets, run as described in Baselines: kernel-predicted lags versus dlmwwbe's fitted lags versus flow-only normalization, evaluated on held-out sheds and held-out healthcare-worker case counts (dlmwwbe).
Experiment three is the gate test on the Swiss Influenza A dataset, where John et al. achieved high-depth sequencing (~836 HA reads per sample versus the sub-200 typical of earlier studies) and validated wastewater clade deconvolution against 9,333 GISAID clinical sequences with R² of 0.71 (clade 6B.1A.5a.2a) and 0.66 (clade 6B.1A.5a.2a.1), including detection of the oseltamivir-resistance mutation S247N (Wastewater sequencing reveals the genomic landscape of Influenza A virus in Switzerland). With clinical clades as the reference standard, we construct ROC curves over concentration-based inclusion thresholds, locate the cost-weighted optimum, and ask whether it detects the resistance mutation and low-frequency clades earlier or at lower sequencing budget than the fixed-gate arm.
The decisive checks, stated plainly: if near-wall deposition physics is real in sewers, lag and attenuation should be predictable from hydraulics that a fitted constant cannot see; and if the Ct < 32 rule of thumb is leaving signal on the table, an optimized gate should find variants that the rule misses. If neither holds, the field learns that the sewer really can be treated as a pipe and the gate really can be a constant — also worth knowing, and currently asserted nowhere with evidence.
References
- Beyond clinical data: the role of sewage monitoring in tracking pandemic trends of SARS-CoV-2
- Numerical simulation of pneumatic conveying characteristics of micron particles in horizontal pipe
- Numerical simulation of flow characteristics in vertical elbows of pneumatic conveying
- dlmwwbe: Dynamic Linear Model for Wastewater-Based Epidemiology
- Estimating incidence of sexually transmissible infections in Australia
- Quadrant asymmetry alteration of retinal degeneration in thyroid-associated ophthalmopathy
- A new model of preoperative systemic inflammatory markers predicting overall survival of osteosarcoma: a multicenter retrospective study
- On the origin and timing of Zika virus introduction in Brazil
- Characterizing influenza A virus lineages and clinically relevant mutations through high-coverage wastewater sequencing