Medical screening AI is usually tuned on one rule: a missed disease is catastrophic, while a false alarm is a cheap inconvenience fixed by a follow-up visit. The rule produces systems like a diabetic retinopathy screener deployed across 45 sites in India. It catches 97% of severe disease, but about half of the patients it refers are healthy. In a well-staffed academic hospital this trade is sensible. In a rural clinic with 1.59 physicians per 1,000 residents, where most diabetic households already spend a crippling share of income on care, the trade looks different. Each false referral uses a specialist slot that a sick patient needed, and it costs a poor family a trip it may not afford. When that happens, the patient may decline the next referral too.
This note proposes measuring a screening system by the number of truly sick patients who actually reach treatment within a clinically safe window, rather than by sensitivity at the moment of the AI's flag. We will build a site-specific referral policy with three outputs: urgent specialist referral, a cheap local re-screen at the patient's next routine visit, and rule-out. The policy is chosen to maximize treated cases given the site's specialist capacity and its patients' likelihood of completing a referral. The claim to test is specific. Where specialist capacity is low and referral is costly, the standard "maximize sensitivity" operating point is past the optimum: lowering false referrals raises the number of sick patients who get care. A middle tier of local re-screening captures most of that gain without the danger of simply raising the threshold.
What exists and what is missing
The asymmetric-cost literature treats a false positive as nearly free. Ami et al. document practitioners who accept roughly 80% false-alarm rates so they never miss a lethal failure, and this attitude motivates high-sensitivity medical triage. The strongest clinical example is Brant et al., who evaluated Google/Verily's retinopathy screener at Aravind sites. They report sensitivity 97.0% and specificity 96.4% for severe or proliferative disease, negative predictive value (the share of cleared patients who are truly healthy) of 99.9%, positive predictive value (the share of referred patients who are truly sick) of 50.7%, and a 0% clinically important miss rate. They use one static operating point for all 45 sites and do not ask what the healthy half of referrals costs patients or specialists. Our work starts from that gap.
Three triage papers provide methods we reuse, and each assumes the downstream human reader has unlimited capacity. Yala et al. set a rule-out threshold so that 19.3% of mammograms skip human reading, with non-inferior sensitivity (90.1% vs 90.6%). Bernstein et al. go further by crediting radiologists for reading more carefully when the remaining batch is enriched. At their safety-first point they report a 36% caseload reduction with zero net missed cancers. At a balanced point they report 75% reduction but 121 net missed cancers. Their accounting of net misses is the closest precedent for our outcome measure, but it is computed at a single well-resourced center with fixed radiologist capacity. Conformal Triage uses distribution-free calibration on local data to split head CTs into guaranteed-negative, guaranteed-positive, and abstain groups. It cuts false positives from 233 (45%) to 8 (5%) while abstaining on 14% of cases. Its abstain group goes to specialists without any limit on queue size. It also needs labeled local calibration data, which low-resource clinics rarely have.
The socioeconomic literature explains why these assumptions break in exactly the settings where screening AI is being deployed. In Gansu, Dang et al. find that 75.19% of diabetic households incur catastrophic health expenditure at a 15% threshold and 35.19% at a 40% threshold, concentrated among the poor. Lin et al. show that rural patients receiving identical radiotherapy at one top cancer center still have a hazard ratio of 3.126 for death. They link this to the physician-density gap between rural and urban areas and to differences in insurance reimbursement. The burden-of-disease analyses, for example Miao et al. and Zhou et al., describe a broader pattern in which detection capacity spreads faster than treatment capacity. All of these are observational. None asks how a detection algorithm should behave once it is deployed into that gap. Our proposal treats these papers as the source of model parameters and plausible ranges, not as evidence of the algorithmic effect.
Chiou et al. call for low-income-country-aware ECG AI but offer no formulation. DR10K and FundusNet improve retinal classifiers under domain shift and label scarcity but evaluate only static metrics. We use them as sources of the score model, not as competitors on the policy question.
Core Claim
Define, for a screening site $s$, the expected number of truly diseased patients treated within a safe window $W$:
$$T_s(\pi) = N_s p_s \sum_{k \in \{\text{refer},\,\text{rescreen}\}} r_k(\pi)\, a_{s,k}\, q_{s,k}(\pi, C_s)$$
In this expression, $\pi$ is the referral policy, meaning the two score cut-offs that define the three tiers. $N_s$ is the screened volume and $p_s$ is the prevalence of disease that threatens sight. $r_k(\pi)$ is the fraction of diseased patients the policy routes to tier $k$, including those caught on re-screen. $a_{s,k}$ is the probability a patient completes that tier's follow-up, which falls with travel and out-of-pocket cost. $q_{s,k}$ is the probability the patient is served within $W$ given specialist capacity $C_s$ and the full referral load, healthy patients included.
False positives appear only through the load term in $q$ and through any effect of past false referrals on $a$. That is the formal sense in which a false positive's cost depends on the site rather than being fixed. The claim is that $\arg\max_\pi T_s$ differs from the maximum-sensitivity policy when $C_s$ and $a_{s,\text{refer}}$ are low, and that the re-screen tier recovers most of the gap. Diabetic retinopathy suits this test because moderate disease usually progresses over months while severe disease scores high, so a delay for mid-score cases is plausibly safe. We will check this assumption rather than take it as given.
We also impose an equity constraint. No site's treated rate for severe or proliferative disease may fall below its rate under the static policy. The goal is to stop false alarms from crowding out the sick, not to ration care for poor sites.
Research Program
We will first obtain a calibrated retinopathy score model, meaning one whose scores can be read as probabilities. It will be pretrained on EyePACS with FundusNet-style contrastive learning and fine-tuned and calibrated on DR10K, an Egyptian telemedicine dataset with 19,378 gradable images. This gives a realistic score distribution under a shift from Western to non-Western populations.
Next we build a discrete-event simulator of a hub-and-spoke screening network. Clinics feed a specialist hub with $C_s$ slots per week. Patients complete referrals with a probability that depends on a cost covariate. We parameterize three regimes, high, middle and low capacity. Physician-density ratios come from Lin et al. (3.92 vs 1.59 per 1,000). The share of patients facing catastrophic cost comes from Dang et al. (35% to 75% depending on threshold). Disease progression during delay is drawn from published ranges and swept widely, since the file contains no progression rates.
Policies are optimized per site over the two cut-offs by grid search on a held-out calibration split. Because realistic adherence and progression parameters are not in any cited dataset, every result is reported as a sensitivity surface over these parameters, not a single number. The scientific product is a map of which regimes favor which policy.
Baselines
The naive reference is the static high-sensitivity point, the Brant et al. deployment style. One threshold is used everywhere and patients are referred first-come-first-served. The strongest reported figures for this approach are sensitivity 97.0%, specificity 96.4% and positive predictive value 50.7% for severe or proliferative disease in Brant et al.. We reproduce the operating point by matching 97% sensitivity on DR10K. We found no reported result on Lacuna for treated-case yield under capacity constraints for any method, so that column starts empty for every baseline.
The second baseline is Youden's J, the threshold that balances sensitivity and specificity. Bernstein et al. use it as their aggressive operating point.
The third baseline is a classical cost-sensitive threshold from Bayes decision theory. It uses a site-specific but fixed cost per false positive, set from local travel cost and income. It captures the "just put poverty into the cost matrix" idea without modeling the queue.
The fourth baseline is Learn-then-Test conformal triage as in Conformal Triage. It is calibrated on the same split, and its abstain group goes to the specialist queue.
The fifth baseline, and the one we consider the strongest competitor, is the static high-sensitivity threshold with the specialist queue ordered by model score rather than arrival time. It keeps every flag but lets sick patients jump ahead of likely-healthy ones.
All policies use the same score model, the same calibration data and the same grid-search budget over their free parameters, and they are evaluated on the same simulated patient streams. Two ablations isolate the proposal's ingredients. The first is a two-tier version of our policy with no re-screen tier, so it can only raise the threshold. The second is our three-tier policy with a fixed false-positive cost replacing the queue model. The proposal beats the field only if it raises severe-disease treated cases above the score-ordered queue in the low-capacity regime, with confidence intervals across simulation seeds excluding zero, while satisfying the per-site equity constraint. It must also reduce false referrals per treated case, our proxy for catastrophic expenditure.
Why not just…?
Why not just set a per-site cost-sensitive threshold? This is the baseline that adds socioeconomic weights to the cost matrix. It treats a false positive's cost as a constant, when under a capacity limit that cost grows with the number of other referrals. A fixed weight therefore cannot tell a site with idle specialists from one with a three-month backlog. The fixed-cost ablation of our own policy tests exactly this. If it matches the full model across the capacity sweep, the queue model is unnecessary and we will say so.
Why not just keep the high-sensitivity threshold and sort the queue by score? This is a real threat. Score ordering removes much of the crowding-out, because likely-healthy patients wait at the back. What it does not remove is the cost to patients: every healthy referred person still travels and pays, and in Dang et al.'s population that is far from harmless. If score ordering matches our policy on treated cases, the remaining argument rests on false referrals per treated case and on adherence. We will report both separately so the result cannot hide behind one aggregate.
Why not just use conformal triage with local calibration? It offers formal guarantees on positive and negative predictive value. But it needs labeled local data, which our target clinics lack. Its abstain group, 14% of cases in the head-CT study, still goes to specialists, which is the scarce resource. We include it and also test it with DR10K calibration transferred to simulated sites, to see whether its guarantees survive. Our re-screen tier can be read as conformal abstention redirected to a cheap local resource. A conformal version of the cut-offs is a natural extension if the basic effect holds.
Why not just train the classifier with a socioeconomically weighted loss? For a calibrated model, reweighting classes in the loss mostly shifts the effective threshold. Any gain should therefore already be available through threshold choice. We will confirm this on DR10K by comparing a reweighted model against the unweighted model at a matched threshold. We do not plan to invest further in training changes unless that comparison surprises us.
Decisive Checks
The first check is whether, in the low-capacity regime, any tested range of adherence and progression parameters makes our three-tier policy beat the score-ordered high-sensitivity queue on severe-disease treated cases. If none does, the conclusion is that the high-sensitivity rule is safe once queues are ordered by score. That negative result would be useful to deployers. The second check is whether the gain comes from the re-screen tier rather than from raising the threshold, which the two-tier ablation settles. The third check is whether the result is driven by the progression assumption. We will report the delay hazard at which the re-screen tier stops paying off, so clinicians can judge whether real retinopathy progression falls on the safe side. The fourth check is whether, as capacity rises toward the high-resource regime, the optimal policy converges back to the Brant-style operating point. That convergence would confirm the simulator reproduces what the academic-center studies found.
References
- "False negative - that one is going to kill you": Understanding Industry Perspectives of Static Analysis based Security Testing
- Performance of a Deep Learning Diabetic Retinopathy Algorithm in India
- A Deep Learning Model to Triage Screening Mammograms: A Simulation Study
- Triage with AI: A Rule-out Framework Quantifying the Risks and Benefits of Screening Mammogram Automation
- Conformal Triage for Medical Imaging AI Deployment
- Factors influencing catastrophic health expenditure of households with people with diabetes in Northwest China-an example from Gansu Province
- Geographical disparities in the prognosis of patients with nasopharyngeal carcinoma treated with intensity-modulated radiation therapy: a large institution-based cohort study from an endemic area
- Global Trends and Regional Differences in the Burden of Infective Endocarditis, 1990–2021: An Analysis of the Global Burden of Disease Study 2021
- The epidemiology of uterine fibroids: global disease burden from 1990 to 2021 and future trend predictions
- Bridging gaps in automated acute myocardial infarction detection between high-income and low-income countries
- DR10K: Transfer Learning Using Weak Labels for Grading Diabetic Retinopathy on DR10K Dataset
- Contrastive learning-based pretraining improves representation and transferability of diabetic retinopathy classification models