Dual-LLM consensus abstraction far more accurate than human abstraction in prostate registry
Study · J Urol · 2026-11-12. Number 1 in the UrologyWatch Top 5 for November 12, 2026.
In 3,210 pathology and 4,739 MRI reports from the Mayo Clinic Prostate Cancer Registry, consensus abstraction using two LLMs achieved median error rates of 0.09% (pathology) and 0.17% (MRI), compared with 2.81% and 1.01% for human abstractors; individual LLMs alone had error rates of 0.27-1.02%. Abstracting all 136,510 data elements took an estimated 32 hours with the LLM pipeline versus 952 hours for human abstraction.
Why it matters: A dual-LLM consensus architecture may substantially cut abstraction time and errors for large clinical registries, supporting consideration for integration into prospectively maintained databases.
Limitation: Findings are from a single-institution registry using specific LLM versions and report formats, so generalizability to other data sources is unknown.
Source: J Urol · PubMed 43013751 · DOI 10.1097/ju.0000000000005387
Loading today's edition…