Rapid diagnostic tools are judged one patient at a time. A cough classifier or chest X-ray model is ranked by its area under the ROC curve (AUC, a threshold-free score of how well it separates sick from healthy). Its decision threshold is then set by a fixed rule, such as balancing sensitivity against specificity. This implicitly assumes that a missed case costs the same whenever it happens. That assumption fails for epidemics with multiple stable outcomes. Recent work on competing infections spreading through group contacts shows that the same disease can either die out or become permanently endemic, depending only on how many people are infected at the start. The line in state space that separates these outcomes is called the basin boundary, or separatrix. Near that line, a few extra missed cases decide the long-run fate of the whole population. Far from it, they barely matter.
We propose to make this dependence usable. We would compute, cheaply, how close a population is to the separatrix, and then choose a screening classifier's operating point (its sensitivity, false-positive rate, speed, and screening rate) to keep the population in the disease-free basin. The core claim is testable. In multistable epidemic models, the classifier that wins on AUC, or the threshold that wins on standard criteria, can lose on the outcome that matters, and which one wins depends on the initial state. If that ranking reversal does not occur for realistic parameters, the idea is wrong. We say below exactly how we would find out.
Core Claim
The dynamics side supplies the phenomenon. Networked competitive bivirus SIS spread with higher order interactions and its conference version Competitive Networked Bivirus SIS spread over Hypergraphs study two viruses competing on a hypergraph, a network whose edges can join groups of three or more people. SIS stands for "susceptible–infected–susceptible": recovered people can be reinfected. These papers prove triple stability: the disease-free state and two single-virus endemic states are all locally stable at once, which cannot happen in pairwise SIS models. Competing spreading dynamics in simplicial complex finds hysteresis and "alternative dominance" governed by the initial infected fraction when triangle-shaped group interactions are added. These papers motivate the work. Neither contains any notion of testing; both treat recovery rates as fixed constants.
The diagnostics side supplies the decision variables. Screening classifiers come with a real trade-off between sensitivity and specificity. The clearest example is COVID-CAPS, a capsule network for chest X-rays. Pre-training on NIH radiographs raised its accuracy to 98.3% and specificity to 98.6%, but dropped sensitivity from 90% to 80%. By AUC (0.97 to 0.99), the pre-trained model is better. Our claim is that in a tri-stable population whose initial infection level lies near the separatrix, the un-pretrained model can be the better public-health tool. In a population far from the boundary, the pre-trained model's lower false-positive load can be better. That reversal only exists because the right answer depends on the initial condition.
A simple observation sharpens the claim. In SIS dynamics, specificity does not affect transmission directly; isolating a healthy person removes a susceptible. Specificity enters only through capacity. False positives consume isolation slots, testing throughput, and compliance. So the operating-point trade-off is really a trade-off among sensitivity, turnaround delay, and capacity, and it has to be modeled as such.
Research Program
We augment the hypergraph bivirus model with screening. Let $x_i^k(t)$ be the probability that node $i$ is infected with virus $k \in \{1,2\}$. Each node is screened at rate $\tau$ by a test with sensitivity $s_k$ for virus $k$; a positive result isolates the person, which adds to their removal rate. The dynamics become
$$ \dot x_i^k = -(\delta_k + \tau s_k)\,x_i^k + \Big(1 - x_i^1 - x_i^2\Big)\Big(\textstyle\sum_j \beta^k_{ij} x_j^k + \sum_{j,l} \beta^{H,k}_{ijl} x_j^k x_l^k\Big). $$
Here $\delta_k$ is the natural recovery rate, $\beta^k_{ij}$ are pairwise transmission rates, and $\beta^{H,k}_{ijl}$ are group (hyperedge) transmission rates. Two further operational features are layered on. The first is a turnaround delay $\Delta$, modeled as an "awaiting result" compartment in which people keep transmitting. The second is an isolation capacity $C$ that caps total positives, true and false, at $\tau\sum_i [s_k x_i^k + f(1 - x_i^1 - x_i^2)] \le C$, where $f$ is the false-positive rate. The pairs $(s_k, f)$ are not free: they are constrained to lie on a real classifier's ROC curve.
The method comes from the monotone-systems literature. Competing Epidemics on Graphs - Global Convergence and Coexistence and Convergence of Bi-Virus Epidemic Models with Non-Linear Rates on Networks show that competitive bivirus models become cooperative under a "southeast" cone ordering. Informally, this orders states so that more of virus 1 and less of virus 2 counts as "larger." Gracy et al. prove strong monotonicity and generic convergence for the hypergraph model. We reuse these results rather than proving new ones.
The practical consequence is that separatrices in strongly monotone systems cannot contain two points ordered with respect to each other. Along any ordered ray of initial conditions, the basin boundary is therefore crossed exactly once, so it can be located by one-dimensional bisection instead of Monte Carlo sweeps. The added testing term $\tau s_k x_i^k$ only changes diagonal terms, so it preserves the cooperative structure. We will prove this; it is a short lemma.
Two features may break monotonicity. The capacity constraint couples nodes through a total and can do so. The delay compartment may or may not preserve it. In those regimes we will verify monotone behavior numerically, or fall back to simulation, and report where bisection stops being valid.
The deliverable is a "basin margin" for each pair of initial state and operating point. It is the signed distance, along the ray from the current state toward the disease-free state, to the separatrix. We then choose the operating point on each classifier's ROC curve that maximizes this margin subject to capacity. A secondary study uses asymmetric sensitivities $s_1 \neq s_2$, as in multiplexed assays like the nanozyme lateral-flow test that detects a bacterium and a virus simultaneously. We ask whether unequal detection shifts the boundary between the two single-virus endemic basins toward the less detectable pathogen.
Because the dynamical results are deterministic, and Doshi et al. list stochastic extensions as open, every prediction is also checked against Markov jump simulations on the same hypergraphs.
Baselines
The naive reference point is no screening: $\tau = 0$. It gives the untreated basin structure and the fraction of initial conditions that go endemic.
The first real competitor is an AUC-ranked, Youden-threshold policy. We deploy whichever classifier has the higher AUC, at the threshold that maximizes sensitivity plus specificity. This is what the diagnostics literature implicitly recommends. The classifier numbers come from reported results. For the acoustic ROC family we use cough-screening models evaluated on the DiCOVA dataset (Coswara recordings):
- The DiCOVA challenge baselines report test AUC 69.85% for cough and 76.85% for breathing.
- The strongest acoustic model in the file, the self-supervised cough model of Harvill et al., reports AUC 85.35 on the blind test set.
We will retrain these models to obtain full ROC curves. The imaging pair uses the two published COVID-CAPS operating points: sensitivity 90% with specificity 95.8%, and sensitivity 80% with specificity 98.6%. Beating this baseline means showing, for a stated set of initial states, a higher probability of reaching the disease-free state with the same classifier pool and the same $\tau$ and $C$.
The second competitor is an effective-reproduction-number rule. It chooses the operating point that minimizes the linearized growth rate at the disease-free state, $\rho(\text{diag}(\delta_k + \tau s_k)^{-1} B^k)$, where $\rho$ is the spectral radius and $B^k$ is the pairwise transmission matrix, under the same capacity constraint. This is the standard epidemiological answer, and it is state-independent.
The strongest competitor for decision quality is brute-force stochastic simulation with grid search. It runs Markov jump simulations for every operating point on a dense ROC grid and every initial state, and picks the best. It will, by construction, be at least as accurate as our method. Our method beats it only by matching its decisions (agreeing on the chosen operating point in at least 90% of scenarios) at an order of magnitude less compute. It will be given the same compute budget cap we report for our method, plus an uncapped "oracle" run.
Two ablations isolate the key ingredient:
- Removing all hyperedges. Pairwise bivirus SIS lacks triple stability, so the reproduction-number rule should become optimal. If basin-awareness still helps there, our story about why it helps is wrong.
- Setting $\Delta = 0$ and $C = \infty$. This checks how much of any reversal comes from delay and capacity rather than multistability itself.
For the dynamical objective, no reported result was found on Lacuna for basin-crossing probability under any testing policy. That comparison has to be generated here.
Why not just…?
Why not just adjust the reproduction number for test sensitivity and turnaround time? In pairwise SIS this is essentially correct, and our first ablation is designed to confirm that. In the hypergraph model of Gracy et al., the disease-free state can be locally stable, with a subcritical linearized growth rate, while an endemic state is also stable. The reproduction-number rule then reports "controlled" regardless of the initial state. It cannot distinguish a small seed from a large one, and hysteresis of the kind found by Pan means the history matters as well. This remains a real threat in one regime. If realistic hyperedge strengths leave the separatrix far from any plausible initial state, the rule will be adequate in practice. We will report the fraction of parameter space where the two rules pick different operating points; if that fraction is negligible, we will say the idea has no practical bite.
Why not just simulate everything with an agent-based model? This is the most serious alternative, and it is our strongest baseline. Simulation is general and needs no monotonicity. Our argument is cost and interpretability. A bisection along ordered rays yields the separatrix location and a margin that can be recomputed as surveillance updates arrive. A grid of stochastic simulations over operating points, initial states, and seeds does not. If the capacity or delay terms break monotonicity in the regimes that matter, this advantage disappears there, and the head-to-head compute comparison will show it.
Why not just always use the most sensitive test, or centralized PCR? A test that is very sensitive but slow lets the infected population grow during the turnaround delay $\Delta$. In a multistable system, that growth can carry the state across the separatrix before results arrive. A test that is very sensitive but has a high false-positive rate saturates isolation capacity. We will sweep $\Delta$ against sensitivity to locate the critical delay beyond which a faster, less sensitive test yields a larger basin margin. If no such crossover exists for realistic capacities, maximizing sensitivity is the right advice, and we will report that.
Why not just use standard cost-sensitive thresholding? That approach sets the threshold from fixed costs for misses and false alarms. Our claim is precisely that the cost of a miss is not fixed: it depends on the distance to the separatrix. If a single fixed cost ratio reproduces the basin-aware choices across initial states, the state-dependence is unnecessary, and this comparison will show it directly.
Proposed Experiments
We use three kinds of network:
- The 4-node hypergraph from Gracy et al., whose triple-stability parameters are known. This serves as a sanity check.
- Random simplicial complexes as in Pan.
- The AS-733 topology used by Doshi et al., with triangles lifted to hyperedges to give a heterogeneous, real network backbone.
Classifier ROC curves come from retrained DiCOVA acoustic models and the two COVID-CAPS operating points. The asymmetric-sensitivity study borrows the idea of unequal per-class detection, visible in VoxMed's uneven per-class F1 on the ICBHI respiratory sound dataset.
The metrics are:
- the probability of reaching the disease-free state under stochastic simulation;
- the deterministic basin margin;
- the agreement rate of chosen operating points with the brute-force oracle;
- wall-clock compute.
Three decisive checks determine whether the proposal survives. First, a ranking reversal must appear. For some realistic initial states, a lower-AUC classifier or non-Youden threshold must beat the AUC/Youden policy on disease-free probability, with confidence intervals from at least 1,000 stochastic runs per cell. The reversal must disappear when hyperedges are removed. Second, bisection along ordered rays must reproduce the Monte Carlo separatrix and the oracle's decisions while using at least ten times less compute. Third, the deterministic margin must predict stochastic outcomes, with calibration improving as network size grows. Where it does not, the finite-size gap is itself a reportable result on the open stochastic question.
References
- Networked competitive bivirus SIS spread with higher order interactions
- Competitive Networked Bivirus SIS Spread Over Hypergraphs
- Competing spreading dynamics in simplicial complex
- COVID-CAPS: A capsule network-based framework for identification of COVID-19 cases from X-ray images
- Competing Epidemics on Graphs - Global Convergence and Coexistence
- Convergence of Bi-Virus Epidemic Models With Non-Linear Rates on Networks - A Monotone Dynamical Systems Approach
- Mercaptophenylboronic Acid‐Mediated Nanozyme Immunochromatographic Assay for Simultaneous Detection of Respiratory Bacteria and Virus
- DiCOVA Challenge: Dataset, Task, and Baseline System for COVID-19 Diagnosis Using Acoustics
- Classification of COVID-19 from Cough Using Autoregressive Predictive Coding Pretraining and Spectral Data Augmentation
- VoxMed: one-step respiratory disease classifier using digital stethoscope sounds