Screening/Referral Rates Compared Across Different Referral Models (Opt-In vs Opt-Out)

What the evidence says

For HPV screening, two referral models are compared directly against a baseline of standard invitation among under- or never-screened individuals: universal mail-out of hrHPV self-sampling kits and an opt-in model where individuals must request a kit. Both increase screening, but at different magnitudes. Universal mail-out produces a relative risk of 2.56 (95% CI 2.10–3.12) and 1534 more people screened per 10,000 (95% CI 1082–2085), drawn from 22 trials totaling 211,031 participants. The opt-in model produces a smaller relative risk of 1.56 (95% CI 1.19–2.03) and 727 more screened per 10,000 (95% CI 247–1338), from a smaller base of 11 trials and 71,433 participants. Both effects carry high confidence in direction and medium confidence in magnitude, and the mail-out estimate rests on roughly triple the trial base of the opt-in estimate.

For diabetic retinopathy referral, AI-assisted screening increases referral uptake compared to standard care, with a relative risk of 1.89 (95% CI 1.18–3.03) across a pooled set of 2 RCTs and 4 observational studies — a 91.9% increase in relative risk and a 23.93% absolute increase. Confidence in the direction of this effect is high, but confidence in its magnitude is low: the pooled estimate carries substantial statistical heterogeneity (I² = 91.9%, elsewhere reported as I² = 92%) and GRADE-assessed low-certainty evidence, with the source citing inconsistent effect sizes across studies and possible publication bias.

That heterogeneity is partly explained by referral-model design. Settings that replaced routine referral (screening everyone who has diabetes) with AI-driven targeted referral showed larger absolute gains — +36.7% to +42.0% — than settings that simply added AI to an existing manual-grading referral workflow, where gains ranged from +11.8% to +23.5% (with a companion claim from the same source reporting +11.8% to +19.4% for this same subgroup — see "Where the evidence splits"). This pattern tracks baseline referral-uptake rate: studies comparing AI against routine-referral pathways started from lower baseline uptake (18.7–22.0%) and showed the largest gains (+36.7% to +42.0%), while studies comparing AI against existing manual-grading workflows started from higher baseline uptake (39.6–77.3%) and showed more modest gains (+11.8% to +19.4%). The source describes this as a clear pattern across all six pooled studies but states that causality is unclear because referral-model type and co-interventions are confounded with baseline rate in the same studies.

No claim in this set addresses an opt-out referral model — one where individuals are automatically enrolled or scheduled by default and must actively decline — as distinct from universal mail-out. The mail-out model tested here removes the request step but does not condition on a default appointment or scheduled visit that the patient must cancel. That distinction is not measured in the current sources.

Where the evidence splits

The two figures for the "AI added to manual-grading workflow" subgroup do not match within the claims themselves: one claim states this subgroup's absolute improvement as +11.8% to +23.5%, while another claim from the same source, quoting the underlying text directly, states +11.8% to +19.4% for the same comparison (studies with baseline uptake of 39.6–77.3%). Both figures come from the same underlying meta-analysis narrative synthesis. The sources do not explain the discrepancy between the 23.5% and 19.4% upper bounds.

Thresholds and dose-response

The evidence on AI-assisted diabetic retinopathy referral establishes a dose-response relationship between baseline referral-uptake rate and the size of improvement from AI assistance. Studies with baseline uptake of 18.7–22.0% (comparing AI to routine referral) show absolute improvements of +36.7% to +42.0%. Studies with baseline uptake of 39.6–77.3% (comparing AI to existing manual-grading workflows) show smaller absolute improvements of +11.8% to +19.4%. The source states this pattern held across all six studies pooled, while noting that referral-model type and co-interventions are confounded with baseline rate and cannot be separated in this data.

Open empirical questions

• Whether an opt-out referral model (default enrollment or scheduling, with the individual required to decline) produces a different screening rate than either the opt-in or universal mail-out models measured here — not addressed in these sources.
• Whether the mail-out vs. opt-in gap seen for HPV self-sampling replicates for other screening modalities, including diabetic retinopathy referral.
• Whether referral-pathway type (routine-to-targeted transformation) or baseline uptake rate is the actual driver of the AI-referral effect size, once co-interventions are disentangled — the source states this is currently confounded.
• Whether increased referral uptake translates into downstream diagnostic confirmation or treatment rates — not measured in any claim here, only the referral/screening step itself.