AI mammogram risk scores climbed for six years before breast cancer diagnosis in a 54,014-woman study

The cancer group's starting score matters more than its final one. Five to six years out, its median was 2.1 ^[8], inside the 1.8 to 2.2 band the cancer-free women held at every time point ^[9][17]. On medians alone, one early reading would not have separated the two groups. The gap opened as the scores moved, and by the index exam the cancer group's median had roughly tripled, to 6.6 ^[8][16]. "The increase in scores among cancer patients was detectable as early as six years prior to diagnosis and became more pronounced over time," said Constance Lehman, the lead researcher ^[10][12].

The 6.6 depends on how the study defined its index exam: the last screening mammogram in the year before diagnosis ^[5]. Of the 817 cancers, 682, or 83%, were detected at screening ^[7]. For most of those women, we think, the index exam is the screen that found the cancer, so the top of the curve partly measures disease already visible on the image. The earlier, smaller rise is the part a risk programme could act on, and the release reports it as group medians ^[8].

Medians describe how two populations moved. Whether a rise in one woman's score predicts her cancer well enough to change her care is a separate question, and a retrospective cohort is where that test begins ^[2]. The pool also narrowed before analysis. Some 89,882 patients were initially included ^[3], and about 40% were excluded ^[18]. Lehman's own wording on clinical value is measured: she said changes in the scores "may provide additional information" ^[15].

The release identifies Lehman as a professor of radiology at Harvard Medical School and chief executive of Clairity, Inc. ^[12]. The model her team ran is described as validated and open-source ^[6]. Her case for image-based risk rests on the women conventional models miss. "This is exciting, because 85% of women diagnosed with breast cancer do not have a significant family history of breast cancer or known genetic mutations," she said ^[11]. The release adds that traditional risk models have limited discriminatory ability in population-based screening ^[13].

The timescales separate cleanly for a screening director. Nothing in this study justifies changing recall or supplemental-imaging rules on a rising score this quarter, because its design was retrospective ^[2]. Whether acting on a trajectory finds cancers earlier than routine screening would is the longer question, and answering it needs a prospective trial.