AIBRIEF
Across 1,613 prostate biopsies at three UK National Health Service hospitals, the same AI tool produced two different kinds of value depending on where it sat in the pathology workflow.
Across 1,613 prostate biopsies at three specialist centres in the UK's National Health Service in England, the same AI tool produced two different kinds of value depending on where it sat in the pathology workflow. The Articulate Pro study, published this month in npj Digital Medicine and led by Associate Professor Clare Verrill at the University of Oxford's Nuffield Department of Surgical Sciences, ran the Paige Prostate AI on 1,049 of those cases.
In concurrent-read mode, where the AI's read reached the pathologist alongside their own, mean turnaround fell 30.1 hours at one site (p < 0.0001). In staged second-read mode, the AI flagged cases after the initial report, and the diagnosis or tumour grade changed in 21 of 386 patients (5.4%). Five of those changes, 1.3% of staged cases, were flagged as potentially affecting clinical management.
Use of immunohistochemistry, the add-on stains that confirm borderline findings, also fell across all three sites, with odds ratios of 0.50, 0.43, and 0.33.
The University of Oxford project page and the Warwick institutional repository copy frame this as prospective clinical evidence rather than a curated lab demo. Industry partner is Paige AI; the York Health Economics Consortium ran the health-economics work, with funding from NHSx and the Accelerative Access Collaborative.
What remains open: the 30.1-hour gain is from one site, the study does not yet say whether the 1.3% management-changing cases were true corrections or new errors, and Paige Prostate's regulatory and reimbursement path outside the UK is unsettled.