AI-Assisted Mammography Reading Finds 17.6% More Cancers in Real-World Use Among 460,000 Women

Most studies of artificial intelligence (AI) for reading mammograms have reanalyzed past images. Nadia Eisemann's team at the Institute of Social Medicine and Epidemiology, University of Lübeck in Germany, published in Nature Medicine the PRAIM study, which analyzed the results of using AI in actual national screening practice.

Observational study of 460,000 women at 12 screening sites

This was a multicenter, observational, real-world non-inferiority study rather than a randomized clinical trial. A total of 463,094 women aged 50 to 69 were screened, and 119 radiologists chose for themselves whether to use AI. As a result, 260,739 women received AI-supported double reading and the rest received standard double reading. The main outcome measures were the cancer detection rate and the recall rate.

The breast cancer detection rate in the AI-supported group was 6.7 per 1,000, which was 17.6% higher than in the control group (5.7 per 1,000) and statistically superior. The recall rate did not increase, at 37.4 vs 38.3 per 1,000, meeting the non-inferiority criterion. The share of recalled women who actually had cancer (positive predictive value) was 17.9% vs 14.9%, and the positive predictive value of biopsy was 64.5% vs 59.2%. In other words, unnecessary recalls, a typical drawback of screening, did not increase.

Implications and limitations

The study offers evidence from large-scale field data that AI can find more cancers without increasing the recall burden. However, because radiologists chose whether to use AI, selection bias related to factors such as image difficulty remains, and the results do not establish causation. Long-term follow-up is needed to learn whether the additional cancers found are overdiagnosis and whether interval cancers, which arise between screenings, and mortality go down.