A boundary set in advance
ACCESS2’s strength is that it did not leave equivalence to a story selected after seeing the results. The peer-reviewed study prespecified a margin for comparing completion of diabetic eye screening. At two Johns Hopkins clinics, 99.5% of the White cohort and 98% of the non-White cohort completed screening. The confidence interval around the 1.5-percentage-point difference lay inside the prespecified margin of plus or minus 10 percentage points. For its measured endpoint, the comparison met its own rule.[1]
The denominator is explicit: of 440 eligible people approached, 397 entered the final analysis and 392 completed screening. The primary comparison concerned completion among these participants. Everyone received the same intervention; this was observational, with no randomized usual-care group. Comparing previous examination rates with current completion makes an access change visible. Assigning the entire difference to AI alone overlooks the free examination and its placement within an existing clinic visit.[1]
The denominator behind perfect sensitivity
The denominator changes when the question becomes diagnostic sensitivity. Among 390 people with images assessable against independent reading-center grading, only eight had disease meeting the referral threshold. The system found all eight, giving 100% sensitivity. Yet the lower bound of its 95% confidence interval was 67.6%. Missing none of eight does not establish perfect sensitivity in another population. A sample that looks substantial for a completion comparison supplies much less information about detecting disease.[1]
Specificity of 76% exposes another burden of the referral decision. Some young people who did not meet the disease threshold also received positive results. The authors discuss reflective structures in young eyes that can affect image interpretation. Screening involves a trade-off between reducing false positives and avoiding missed disease. Independent expert readers masked to the AI output provide a strong comparison here; the small disease denominator and unnecessary referral burden constrain different parts of that comparison.[1]
Young people outside the clinic
The main boundary on transferring the access finding to all young people begins at the clinic door. Participants were already attending diabetes care, and screening was free. Some had experience with similar screening. The behavior of people who never reached care was not measured in this sample. Similar completion across the groups is a real achievement of this implementation. Restricting its scope describes where that achievement occurred rather than erasing it.[1]
The value of a subsequent comparison depends on the question defined in advance: does completion persist in routine paid care, can services reach people outside care, and is necessary assessment completed after referral? These are different links in the chain toward preventing vision loss. ACCESS2 passed a prespecified test at the screening-completion link. Diagnostic accuracy and population-wide access require more disease cases and different service conditions. Placing eight cases beside the perfect-sensitivity figure makes clear which question the study answers strongly.[1]