In a retrospective study of 88,963 mammograms, an AI risk model flagged roughly 20 percent of breast cancers up to six years before they were formally diagnosed, at 90 percent specificity. The work was led by Fredrik Strand and colleagues.