PRAIM: What the largest prospective AI study in healthcare means for breast screening
For years, the case for AI in radiology rested on retrospective data and small, controlled settings. That changed on 7 January 2025, when Nature Medicine published the PRAIM study — the first nationwide deployment of AI in Germany’s National Screening Program.
PRAIM enrolled 500,000 participants across Germany. It was conducted by Vara and the University of Lübeck. It is the largest prospective, real-world study of AI in healthcare to date.
What PRAIM measured
The study integrated AI into live screening workflows and tracked what happened to two things: cancer detection and radiologist workload.
The findings:
AI integration improved cancer detection rates by 17.6%, without increasing false positives
Radiologists spent 43% less time interpreting AI-tagged normal examinations
In a simulated scenario where radiologists did not interpret AI-labelled normal mammograms — 56.7% of all scans — cancer detection rate still improved by 16.7%, while recalls fell by 15%
These are not lab results. They come from a live national screening programme, across multiple centres, with no exclusion criteria.
Why this study is different
Previous AI studies in radiology have typically relied on retrospective datasets or controlled research environments. PRAIM was prospective and deployed in routine clinical practice — the same conditions under which AI would need to perform if adopted at scale.
That distinction matters. Retrospective studies can show that AI performs well on historical data. Prospective studies show whether it performs well when radiologists actually use it, in real workflows, on real patients.
Professor Alexander Katalinic, Head of the Institute of Social Medicine and Epidemiology at the University of Lübeck and lead investigator:
“The results are extremely positive and have exceeded our expectations. We can now demonstrate that AI significantly improves the cancer detection rate in screening for breast cancer.”
What this means in practice
Germany screens millions of women annually. If PRAIM’s results were applied across the full programme, the implications are straightforward: more cancers caught earlier, fewer unnecessary recalls, and a meaningful reduction in the reading workload that screening programmes are struggling to staff.
The 56.7% workload finding is particularly relevant. More than half of all scans were AI-labelled normal — and even without a radiologist reviewing those, detection rates still improved. For programmes facing chronic radiologist shortages, that is not a theoretical efficiency gain. It is a practical path forward.
Context
Vara’s AI is integrated into 40% of Germany’s national screening centres, processing over 80,000 mammograms monthly. The company has published in Nature Medicine and Lancet Digital Health, and holds what is now the largest prospective evidence base for AI in any radiology specialty.
Stefan Bunk, CTO of Vara:
“The PRAIM study is a testament to the vast potential of AI in improving breast cancer screening. We are confident that the results will elevate the discussion about AI-assisted mammography screening, paving the way for wider adoption in health systems globally.”
PRAIM does not close the conversation about AI in screening. But it shifts it from “does AI work?” to “how do we deploy it responsibly?” — a question that requires a different kind of evidence, and a different kind of infrastructure.