An FDA-approved AI model applied to prior-year screening mammograms in 1,509 women across four U.S. states showed modest ability to predict biopsy-confirmed cancer (AUC 0.62), and among 508 cancers, higher prior-year scores were linked to lower-grade tumors. In this retrospective study, Grade 3 tumors had lower prior-year AI risk scores than Grade 1 tumors, suggesting the model may detect subtle imaging features of low-grade malignancy before clinical detection.

Why It Matters To Your Practice

• AI risk scores from earlier mammograms may reflect biologic behavior, not just near-term cancer presence.

• If validated, these tools could help clinicians better understand which cancers are more likely to leave subtle imaging signals a year before diagnosis.

• The study also adds explainability: higher AI scores may be more associated with low-grade disease than with more aggressive tumors.

Clinical Implications

• Do not treat these scores as stand-alone cancer detectors: discrimination was modest, with an AUC of 0.62.

• Lower scores may not reassure against higher-grade disease, since Grade 3 tumors trended toward lower prior-year risk scores.

• For breast imagers and referring clinicians, AI outputs may eventually be more useful for contextual risk stratification than for binary rule-in/rule-out decisions.

Insights

• The cohort included women biopsied after a 2021 screening mammogram; 33.7% had biopsy-confirmed malignant neoplasm.

• In univariate analysis, invasive lobular carcinoma showed higher prior-year scores than ductal carcinoma in situ, but that association was no longer significant after adjustment for grade.

• This suggests tumor grade may mediate some of what the AI model is detecting on prior mammograms.

The Bottom Line

• Prior-year mammography AI risk scores may pick up faint imaging patterns linked to low-grade breast cancer about 1 year before diagnosis.

• That is promising for understanding model behavior, but the current performance is not strong enough to support overreliance in practice.

• For now, think of this as an explainability and hypothesis-generating signal, not a practice-changing screening endpoint.