Impact of Artificial Intelligence on Breast Cancer Screening

The goal of this clinical trial is to compare patient-centered outcomes when 3D screening mammograms are interpreted with versus without a leading FDA-cleared AI decision-support tool in real-world U.S. settings and to assess patient's perspectives on AI in medicine.

The main questions it aims to answer are:

1. Will AI use be associated with an increase in cancer detection and an initially higher recall rate as radiologists start using AI, followed by a recall rate comparable to that without AI (no more than 1.5 percentage-points higher) after a learning curve period? Will AI use will be associated with lower rates of missed breast cancers and similar rates of false alarms after a learning curve period?
2. Will improved patient outcomes with AI be most pronounced for exams on women who are White, older, and have less dense breasts, and on baseline exams? Will AI aid patient outcomes when the interpretation is by radiologists with less clinical experience, lower annual interpretive volume, and less tolerance of ambiguity? Yet, will there be greater automation bias (the tendency for humans to defer to a computer algorithms' results) noted among these radiologists?
3. What are patients' perspectives on AI in mammography, including their confidence in breast cancer screening when interpreted with vs. without AI? What are patients' perspectives on the importance of the study results?

Researchers will compare patient-centered outcomes when 3D screening mammograms are interpreted with versus without a leading FDA-cleared AI decision-support tool in real-world U.S. settings.

This trial will include all adult patients undergoing 3D mammography breast cancer screening at imaging facilities across University of California at Los Angeles and University of Washington health systems and all radiologists interpreting breast cancer screening. All screening mammograms at these facilities will be randomized to either intervention (radiologist with AI support) versus usual care (radiologist alone) to see if interpreting these mammograms with the AI tool's assistance improves patient outcomes.

AIM 1. Conduct a 2-year randomized controlled trial (RCT) and assess immediate performance measures and outcomes of the diagnostic evaluation cascade for 3D screening mammograms interpreted with vs. without AI: Outcomes will include recall rate, cancer detection rate, and positive predictive value (PPV1), along with measures of the diagnostic evaluation cascade including short-interval follow-up rate, biopsy rate, and biopsy yield (i.e., PPV2 and PPV3).

AIM 2: Assess 1-year breast cancer performance measures and clinical patient outcomes for 3D screening mammograms interpreted with vs. without AI: After linkage with state and regional tumor registries, we will compare longer-term performance measures (sensitivity, specificity, false positive rates) and patient outcomes (interval cancer rates and tumor molecular subtypes) associated with screening with vs. without AI.

AIM 3: We will perform subgroup analyses of the interaction of patient-, exam-, and radiologist-level characteristics associated with improved screening performance with AI. Patient characteristics will include demographics (e.g., age, race/ethnicity) and risk factors (e.g., breast density, family history of breast cancer, previous breast biopsy, prior breast cancer). Exam-level factors will include baseline vs. subsequent screening exams and mammography unit manufacturer. Radiologist characteristics will include experience level (e.g., general clinical experience and experience with AI), tolerance of ambiguity in clinical decision-making and trust in AI, measured with standardized survey tools.