July 2026 — Doxiverse Staff
A prospective, multi-year study across three NHS hospital trusts found that AI-assisted review of prostate biopsies led pathologists to change a diagnosis or cancer Grade Group in 5.4% of cases evaluated during a staged second-read phase. Changes in 1.3% of that reviewed cohort were considered potentially significant enough to affect the clinical management options available to patients.
The peer-reviewed study was published in npj Digital Medicine and was highlighted by the University of Oxford in July 2026.
Why it matters: This is prospective, peer-reviewed, multi-site evidence from routine NHS practice rather than a vendor white paper. It provides real-world evidence that AI-assisted pathology can influence diagnostic decisions while also potentially improving workflow through faster turnaround times and fewer requests for additional staining.
The study was publicly funded and academically led, although Paige participated as the industry partner and contributed to study design and manuscript preparation. The authors state that Paige was not involved in the statistical analysis or interpretation of the study findings.
What Happened
The study, called Articulate Pro, ran for roughly two years across Oxford University Hospitals NHS Foundation Trust, North Bristol NHS Trust, and University Hospitals Coventry and Warwickshire NHS Trust. It evaluated the Paige Prostate AI system within routine prostate biopsy reporting workflows. Paige was acquired by Tempus AI in 2025 and is now part of Tempus' diagnostics business.
In total, 1,613 prostate biopsy cases were included in the study, with 1,049 reported using AI assistance in some form.
The most clinically notable diagnostic-change results came from a smaller staged second-read cohort. In this phase, pathologists first reviewed the biopsy and recorded their intended diagnosis before seeing the AI output. They then reviewed the AI findings and could revise the diagnosis before finalizing the report.
Among 386 cases evaluated through this staged process, pathologists changed the diagnosis or cancer Grade Group in 21 cases, or 5.4%.
A hospital discrepancy review process subsequently determined that changes in five of the 386 cases, or 1.3%, could have been clinically significant enough to alter the management options available to the patient.
The staged second-read analysis was completed at two of the study's three sites. The third site was unable to complete this phase because of logistical difficulties associated with the additional data entry and adjudication required.
Among the clinically significant cases were two biopsies initially intended to be reported as benign. In one, AI-assisted review contributed to identification of Gleason 3+3, Grade Group 1 prostate adenocarcinoma, after which the patient entered active surveillance. In another, the diagnosis was revised to atypical small acinar proliferation, or ASAP, resulting in continued PSA surveillance.
Other changes involved alterations in cancer Grade Group that could affect whether options such as active surveillance or more definitive treatment would be offered.
Importantly, the study evaluated changes in diagnostic interpretation and available management options; it did not demonstrate that AI ultimately changed the treatment received by every patient whose diagnosis was revised.
Workflow Effects
The study also examined whether AI assistance affected pathology workflow.
At one study site, AI-assisted reporting was associated with an adjusted reduction in average total turnaround time of approximately 30 hours. A similar numerical reduction was observed at another site but did not reach statistical significance, while no significant improvement was demonstrated at the remaining site.
AI assistance was also associated with significantly fewer requests for additional immunohistochemistry staining at all three hospitals. The magnitude of the reduction varied between sites, with adjusted odds ratios ranging from approximately 0.33 to 0.50.
For pathology laboratories dealing with increasing biopsy volumes, staffing constraints, and pressure to shorten reporting times, these workflow findings may ultimately be as important as the diagnostic-change results.
Professor Clare Verrill of the University of Oxford's Nuffield Department of Surgical Sciences, who led the research, said the findings provide evidence that AI can be deployed safely and effectively alongside expert pathologists to support prostate cancer diagnosis within the NHS.
What This Means for Pathology Departments and Hospitals
Articulate Pro is notable because it evaluated AI prospectively within routine clinical workflows rather than relying only on retrospective datasets or comparisons against a previously established reference diagnosis.
That distinction matters.
Many pathology AI studies focus primarily on measures such as sensitivity, specificity, concordance, or area under the receiver operating characteristic curve. Those metrics are important, but they do not necessarily show what happens when a practicing pathologist actually uses the software while reporting real patient cases.
The staged second-read component of Articulate Pro provides a more practical measure: how often did seeing the AI output cause the pathologist to change what they were planning to report?
In this cohort, that happened in 5.4% of cases.
The smaller 1.3% figure provides another layer of clinical context by identifying cases where those diagnostic changes could potentially have affected the patient's management options.
For laboratory administrators evaluating AI pathology platforms, the reductions in turnaround time and additional immunohistochemistry may be particularly relevant when considering operational efficiency and staffing pressures.
For pathologists, urologists, and oncologists, the diagnostic and grading changes may be more important when considering whether AI functions as a useful second set of eyes during biopsy interpretation.
Results Could Differ Outside Specialist Centers
Another important limitation is the setting in which the study was performed.
The participating hospitals were academic centers where prostate biopsies were interpreted by experienced specialist pathologists.
The authors note that the effect of AI could potentially be greater in non-specialist environments. Previous research has suggested that general pathologists may experience larger improvements from AI assistance than specialist uropathologists.
That does not mean the 5.4% or 1.3% findings can simply be extrapolated to a community hospital. Results in non-specialist laboratories could differ in either direction.
Instead, it highlights an important unanswered question: whether AI-assisted pathology provides greater incremental benefit when baseline subspecialty expertise is less readily available.
What's Still Unsettled
Several limitations are worth considering before applying these findings more broadly, particularly in the United States.
- The 5.4% and 1.3% figures do not come from the entire 1,613-case study population. They come from the 386 cases that underwent the staged second-read process at two of the three participating sites.
- The study was not a randomized controlled trial. It was a registered, prospective, multicenter observational service evaluation conducted within NHS clinical workflows.
- The 1.3% figure represents cases in which the diagnostic change could potentially affect clinical management options. It should not be interpreted as evidence that AI definitively changed the treatment ultimately received by 1.3% of patients.
- Long-term patient outcomes were not assessed. The study was not designed to determine whether AI assistance reduced metastases, prostate cancer mortality, overall mortality, or other long-term clinical outcomes.
- This paper did not report a formal cost-effectiveness analysis. A separate economic analysis has been conducted as part of the project and is expected to be reported separately.
- Paige was involved in the project. The company participated as the study's industry partner and contributed to study design and manuscript preparation. Paige-affiliated authors were included on the publication. The authors state, however, that the company did not conduct or influence the statistical analysis or interpretation of the findings.
- The study was conducted within the UK NHS and its regulatory and clinical framework. US pathology departments would need to evaluate the technology according to the specific FDA-authorized indication applicable to the product and workflow being considered.
A Note on US FDA Authorization
Paige Prostate Detect previously received FDA De Novo authorization as an adjunctive second-read tool for prostate biopsy interpretation.
Its FDA-authorized role is to assist pathologists after their initial review by identifying areas suspicious for prostate cancer that may warrant additional examination. The AI output is not intended to replace the pathologist or serve as the primary diagnosis.
The broader Paige Prostate platform evaluated within the UK Articulate Pro program included functionality beyond simple cancer detection. The FDA authorization should therefore not be interpreted as blanket authorization for every AI function evaluated during the NHS study.
US pathology practices considering similar technology should review the specific FDA-authorized indications, intended use, and workflow requirements for the product they plan to deploy.
The Bottom Line
Articulate Pro provides relatively strong real-world evidence that AI can have a measurable effect when incorporated into routine prostate biopsy reporting.
During the staged second-read phase, AI-assisted review led pathologists to change a diagnosis or cancer Grade Group in 5.4% of reviewed cases. Changes in 1.3% of that cohort were considered potentially significant enough to affect the patient's available management options.
At the same time, AI assistance was associated with fewer requests for additional immunohistochemistry across all three study sites and a substantial reduction in turnaround time at one center.
The study does not prove that AI improves long-term prostate cancer outcomes, nor does it establish the economic case for widespread deployment. Its findings also come primarily from specialist academic pathology environments.
But it addresses an increasingly important question for healthcare AI: not simply whether an algorithm performs well against a benchmark, but whether introducing it into real clinical practice actually changes what clinicians do.
In this study, sometimes it did.
Sources:
- Browning, L. et al. “An evaluation of artificial intelligence assisted prostate biopsy reporting in the Articulate Pro study.” npj Digital Medicine 9, 537 (2026).
https://www.nature.com/articles/s41746-026-02592-8 - University of Oxford. “Study shows AI can be safely deployed in NHS prostate cancer diagnosis.” July 2026.
https://www.ox.ac.uk/news/2026-07-14-study-shows-ai-can-be-safely-deployed-in-nhs-prostate-cancer-diagnosis - U.S. Food and Drug Administration. Paige Prostate De Novo classification, DEN200080.
This article summarizes a UK-based clinical study and its regulatory context. It is not a substitute for reviewing the applicable FDA authorization, product labeling, or institutional requirements for any AI pathology system being considered for clinical use in the United States.