clinicians.build · interactive · September 26, 2026 · built on Graham Walker, MD, in Every

A Thousand Mammograms

“AI makes uncertainty cheap to explore without making it cheap to resolve,” writes Graham Walker. A screening callback is what resolving looks like: a phone call, a second visit, sometimes a biopsy. Below are 1,000 women screened. The cancers stay the same no matter where you move the slider. What changes is how many women get the call.

Primary source: Graham Walker, MD, thesis statement, Every.
Data: hospital recall rates from CMS Care Compare Outpatient Imaging Efficiency – Hospital, OP-39, Medicare outpatients, Jul 2024–Jun 2025 (… hospitals). Cancers per 1,000: Breast Cancer Surveillance Consortium benchmarks (Lehman et al., Radiology 2017: 5.1 detected, 0.8 missed per 1,000 screens).

Read Graham Walker’s essay → Or read today’s newsletter →

…

Recall rate —

cancer foundcancer the mammogram missedcalled back, no cancerall clear

Below 5%: the guideline’s lower edge. Benchmarks suggest that holding back this hard starts to cost sensitivity. This model keeps cancers fixed, so it can’t show that cost. Treat this end of the slider as optimistic.

Called back–

…without cancer–

A callback is cancer–

Hospitals recalling more–

Strip: every hospital’s actual OP-39 rate (one tick each). The red line is your slider.

The part that doesn’t move About six women in a thousand have a cancer on this mammogram, and a good reader finds about five of them. Go from the median hospital (…) to the 90th percentile (…) and you make roughly … more phone calls per thousand. Under this model the extra calls find no extra cancers, because the median reader is already inside the guideline window. That model is simplified on purpose, and the caveats below say where. Still, this is Walker’s point in numbers. Each extra callback took a reader one moment of doubt to order, and the patient carries it until the follow-up comes back.

Stress-test the far right of the strip The hospitals past 20% sit mostly in Connecticut, New York, New Jersey and Florida. There, a same-day supplemental ultrasound for dense breasts probably counts as a “recall” under OP-39. For some of those women the second test may have been planned before anyone read the first image. If so, that’s a policy decision, not a reader’s. The explorer lets you remove those states.

Where this is thin — read before quoting

The cancers are a benchmark, not this data. CMS publishes recall rates only. Cancer counts come from BCSC community practice, 2007–2013, all ages. Medicare-age women have higher cancer rates, so the fixed dots undercount them somewhat.
Fixed sensitivity is a simplification. In real practice, readers who recall very little miss more, and readers who recall more catch slightly more, with fast-diminishing returns. The waffle holds detection constant so that one variable moves.
Claims-based. OP-39 counts any breast imaging within 45 days of a screen, and it correlates only moderately with radiologists’ own registry numbers (r = 0.43).