Pooled testing saves lab work by mixing samples from several people and testing the mixture once. Modern assays report more than "positive or negative." A qPCR machine (the standard PCR test that measures how many amplification cycles, the Ct value, it takes for viral signal to appear) reports how much virus is in the pool. Labs usually coarsen that reading into a few bins, such as low, medium, or high risk. Today the cutoffs between those bins are chosen either to maximize information or by heuristics about assay noise. Separately, the decision about who gets an individual retest comes from fixed rules attached to the bins. This note proposes choosing the bin cutoffs, the pool size, and the retest rule together, so as to minimize the real operating cost: tests spent plus the cost of sending an infected person home. It also requires that the per-person infection probabilities the decoder produces be trustworthy enough for a cost-optimal threshold to mean anything.
The motivation is a mismatch visible across three literatures. Group testing theory shows that a few output levels recover most of the value of exact counts. Decision theory for "reject options" shows that the best abstention threshold is a closed-form function of costs, but only when probabilities are calibrated. Work on low-precision neural networks shows that coarsening numbers distorts confidence in exactly the range where thresholds are set. No one has asked which quantizer is best when its output feeds a cost-sensitive decision to accept or defer, rather than an information count. We will answer that on real clinical data.
Core Claim
The setting is a quantized adder channel: each pool's reading is the sum of its members' signals, blurred by noise and dilution, then mapped into one of $L$ bins. The claim is that the bin cutoffs that minimize expected operational cost differ measurably from both the information-maximizing cutoffs and the empirical cutoffs used in practice. For $L=3$, choosing them jointly with pool size and a Chow-style retest rule should lower cost per person at matched false-negative rate, compared with the published semi-quantitative protocol.
Concretely, let $\theta$ be the bin cutoffs on pooled Ct, $g$ the pool size, and $p_i$ the decoder's posterior probability that person $i$ is infected given all pooled bin readings. After the first round each person is cleared, declared positive, or deferred to an individual retest. The design objective is
$$J(\theta, g) = \mathbb{E}\Big[\tfrac{T(\theta,g)}{N}\,c_{\text{test}} + c_{\text{miss}}\,\mathbb{1}[\text{cleared} \wedge \text{infected}] + c_{\text{fa}}\,\mathbb{1}[\text{declared} \wedge \text{healthy}]\Big],$$
where $T$ is the total number of tests, $N$ the cohort size, and the $c$'s are per-event costs. For fixed $(\theta,g)$ and calibrated $p_i$, the best retest rule is a pair of thresholds on $p_i$. These follow from the formula in Rethinking and recomputing the value of machine learning models, which derives the risk-minimizing confidence cutoff $\tau = (C_{\text{error}}-C_{\text{reject}})/(C_{\text{error}}-C_{\text{correct}})$, with retest cost playing the role of rejection cost. Deferral is worthwhile only under the sandwich condition surveyed in Machine learning with a reject option: a survey. The decision layer therefore reduces to a closed form, and $J$ becomes a function of $(\theta,g)$ alone. Optimizing $J$ over $(\theta,g)$ is the new part.
We expect the cost-optimal cutoffs to differ from the information-optimal ones for a specific reason. Amin Emad found by numerical search that capacity-maximizing quantizers want a very narrow bin isolating sums of one or two defectives. Under a cost objective with large $c_{\text{miss}}$, however, what matters is resolving the Ct region where one weak positive, diluted by a factor $g$, sits near the assay limit. There the clear-or-retest decision flips. That region is set by the dilution law $Ct_{group}\approx Ct_{individual}+M\log_{10}(g)$ from Nambiar et al., so the optimal cutoffs should move with $g$. This is a hypothesis the experiments test, not an assumption.
Relation to Prior Work
Two papers supply the channel and the data. Amin Emad formalized semi-quantitative group testing as an adder channel followed by a quantizer, and optimized thresholds for capacity. We keep the channel but replace capacity with decision cost. Semi-quantitative group testing for efficient and accurate qPCR screening built a two-cutoff, three-bin protocol with dual-permutation pooling on 133,816 Israeli SARS-CoV-2 tests. It reports 24% fewer tests than standard pooling and an optimal pool size of 15 versus Dorfman's 8 at 2% prevalence. Its cutoffs are fixed from assay noise, and its second stage retests by bin membership. We reuse its data, dilution model, and pooling design, and change how cutoffs and retests are chosen. The two-round result in Semiquantitative Group Testing in at Most Two Rounds constrains us to two stages, which is also what labs run.
The decision layer comes from the reject-option literature. On the Calibration of Multiclass Classification with Rejection proved that decoupled training followed by post-hoc thresholding of the top probability is Bayes-consistent, while joint classifier-rejector losses are not for more than two classes. That result is why we threshold the decoder's posterior post hoc rather than learning a rejector.
That result also marks our departure. Decoupling is optimal given the full input. Here the quantizer discards information before the decoder ever sees it, so the choice of quantizer cannot itself be decoupled from the downstream cost. A Connection Between Learning to Reject and Bhattacharyya Divergences gives a more conservative rejection rule, but it assumes an oracle posterior. We test it only as an alternative decision layer.
The calibration concern comes from quantized neural networks. When Quantization Affects Confidence of Large Language Models? found that 4-bit quantization raises calibration error (about 10% on HellaSwag for LLaMA-7B) while overconfidence on errors stays near 0.95. Precision-Induced Miscalibration shows that confidence distortion peaks in the 0.7–0.9 range where thresholds are usually placed. Its authors state that their continuous correction breaks down in coarse discrete regimes.
A three-bin quantizer is exactly such a regime. The decoder's posterior is computed through a misspecified dilution-plus-noise model, so we expect analogous miscalibration and must measure and correct it before trusting $\tau$. These papers motivate a calibration check. We borrow no method from them, because their corrections target floating-point roundoff, not bin coarsening.
Research Program
We will build a pooling simulator grounded in the Israeli dataset's individual follow-up results. Pools are formed by resampling individuals at prevalence 2%, 5%, and 10%, matching the simulated cohorts in Nambiar et al.. Pooled Ct is generated from individual Ct through the dilution law plus an empirical noise distribution. We fit that noise distribution to the dataset's real pooled readings and hold out part of those readings to validate the simulator.
On this simulator we will implement a belief-propagation decoder that returns per-person posteriors from binned pooled readings. We will measure the decoder's calibration against follow-up labels and, if needed, correct it with a monotone recalibration fitted on held-out pools. For each candidate $(\theta, g)$, we then apply the two Chow thresholds and estimate $J$ by Monte Carlo.
The search space is small ($L\le 4$, one integer $g$), so we optimize by exhaustive grid plus local refinement. The contribution is the objective, not a clever optimizer. We will sweep the cost ratio $c_{\text{miss}}/c_{\text{test}}$ across two orders of magnitude and trace how the optimal cutoffs move. That tells us whether the decision-critical Ct band shifts with $g$ as the dilution argument predicts.
Baselines
The naive reference point is classical Dorfman pooling: binary pool results with individual retests of every positive pool, run at its own optimal pool size on the same resampled cohorts. The strongest existing method is the three-bin protocol of Nambiar et al., run with its published cutoffs, permutation design, and bin-based retest rule on the identical simulator and cohorts. Its reported figure is 24% fewer tests than standard pooling on the Israeli data. No false-negative rate paired with that figure was found on Lacuna. We will therefore recompute its test count and false-negative rate ourselves, and require that any win be at matched or lower false-negative rate.
A third baseline uses the capacity-optimal cutoffs in the style of Amin Emad, found by the same grid search with mutual information as the objective. This isolates whether the cost objective, rather than simply having optimized cutoffs, drives any gain.
Two ablations isolate the key ingredient. One keeps the published cutoffs and adds only our calibrated Chow retest rule, so it tests the decision layer alone. The other optimizes cutoffs for cost but retests by bin membership, so it tests the quantizer alone.
As a ceiling we run the same decoder on unbinned Ct. Every method gets the same cohorts, decoder, and noise model. Every tunable quantity gets the same grid budget, tuned on a separate set of resampled cohorts from the ones used for evaluation. Beating the strongest baseline means lower $J$ with confidence intervals excluding zero, over resampled cohorts, across most of the cost-ratio sweep.
Why not just…?
Why not just keep the published cutoffs and add a cost-optimal threshold on the decoder's posterior? This is the most serious threat. The decoupling result of Ni et al. suggests post-hoc thresholding should be enough. It is precisely the first ablation above. If it matches the full joint design within noise, we will report that the gain comes from the decision layer, and the quantizer-design claim fails. We think it will fall short because the posterior can only be as sharp as the bins allow. If no cutoff lands near the diluted-weak-positive band, no threshold on $p_i$ can separate those people from healthy pool-mates.
Why not just decode the raw continuous Ct values? Where continuous readings are reliable, this may dominate, and we include it as a ceiling rather than argue against it. The paper's contribution is then measured as the fraction of the gap between published three-bin and continuous decoding that cost-aware binning closes at $L=2,3,4$. That gap matters for coarse point-of-care readouts and for labs where continuous Ct noise models do not transfer across machines. We will also test continuous decoding under a deliberately misspecified dilution slope $M$ to see whether binning is more robust.
Why not just recalibrate with temperature scaling and keep everything else? A single temperature cannot move bin boundaries, so it cannot recover information the quantizer discarded. Precision-Induced Miscalibration also reports that one global temperature leaves input-dependent distortion uncorrected. We include temperature-scaled posteriors as one recalibration option inside the decision layer. If it suffices for calibration, we adopt it, since calibration is a prerequisite here, not the claim.
Decisive Checks
Three measurements settle the proposal. First, per-person posterior calibration error, measured against individual follow-up results, must be low after recalibration. Otherwise the Chow thresholds are not cost-optimal and the comparison is uninterpretable. Second, across the cost-ratio sweep, the cost-optimized quantizer must beat both the published protocol and the capacity-optimal quantizer in expected cost at matched false-negative rate, and the decision-layer-only ablation must not close that gap. Third, the optimal cutoff locations should move with pool size, tracking the diluted-weak-positive band predicted by $M\log_{10}(g)$. If the gains appear but the cutoffs do not move this way, the dilution mechanism is wrong even if the method works, and we will say so.
References
- Rethinking and recomputing the value of machine learning models
- Machine learning with a reject option: a survey
- Semiquantitative Group Testing
- Semi-quantitative group testing for efficient and accurate qPCR screening of pathogens with a wide range of loads
- Semiquantitative Group Testing in at Most Two Rounds
- On the Calibration of Multiclass Classification with Rejection
- A Connection Between Learning to Reject and Bhattacharyya Divergences
- When Quantization Affects Confidence of Large Language Models?
- Precision-Induced Miscalibration: Understanding and Correcting Confidence Distortion in Quantized Neural Networks