Multicenter benchmarking of large language models for bladder cancer management recommendations from mpMRI reports

Large language models (LLMs) can interpret radiology reports, but their ability to generate guideline-concordant recommendations from bladder cancer multiparametric MRI (mpMRI) reports remains uncertain. Following TRIPOD-LLM, we retrospectively evaluated nine API-accessed LLMs using 590 mpMRI reports from three hospitals under two input conditions: original Chinese reports and clinician-translated, quality-controlled English reports. Two physicians independently rated each model-generated recommendation on a 1–5 ordinal scale. We analyzed 21,240 ratings using patient-level random-intercept cumulative link mixed models, with paired Wilcoxon tests as secondary exploratory analyses. The association between input condition and physician rating differed across LLMs (likelihood-ratio chi-square = 683.05, df = 8, P < 0.001). Compared with original Chinese input, clinician-translated English input was associated with higher ratings for chatgpt-o3-pro (OR 2.79, 95% CI 2.29–3.41), deepseek-v4-flash (OR 1.46, 95% CI 1.19–1.81), claude-opus-4-7 (OR 3.07, 95% CI 2.47–3.82), and gpt-5.5 (OR 1.88, 95% CI 1.53–2.32), but with lower ratings for claude-sonnet-4-5 (OR 0.29, 95% CI 0.23–0.36; all Holm-adjusted P < 0.001). Four LLMs showed no clear input-condition difference. Center-specific contrasts also depended on input condition and were materially unchanged after adjustment for age, sex, pathological T stage, tumor grade, tumor number, and maximum lesion diameter. Inter-rater agreement was substantial (quadratic weighted kappa range, 0.621–0.743). Readability did not track physician-rated recommendation quality. These findings argue against clinician-translated English as a universal optimization strategy and support model-specific validation of input conditions before clinical deployment. LLM-generated recommendations from mpMRI reports are best positioned as clinician-supervised, imaging-based decision support rather than standalone treatment advice.

Authors

Institutions

Publication Details

Journal
npj Digital Medicine
Published
2026-11-13
DOI
https://doi.org/10.1038/s41746-026-03469-6
Primary Topic
Artificial Intelligence in Healthcare and Education
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Multicenter benchmarking of large language models for bladder cancer management recommendations from mpMRI reports

Zhenghao Liu, Yaxin Hou, Hongjie Shi, Yingchun Kuang et al.
7 links
npj Digital Medicine
Artificial Intelligence in Healthcare and Education
article

Multicenter benchmarking of large language models for bladder cancer management recommendations from mpMRI reports

Zhenghao Liu, Yaxin Hou, Hongjie Shi, Yingchun Kuang, Lilong Liu, Ke Chen, Zongyu Li, Yang Li, Pengjie Shi, Chong Shen, Chengzhang Zhang, Qinliang Si, Guang Zeng, Zheng Liu, Yuhong Ding, Yiting Liu, Xiaoni Chen, Junyi Hu, Hailong Hu, Yixian Li, Liangqi Kang, Lina Yu
article en

Abstract

Large language models (LLMs) can interpret radiology reports, but their ability to generate guideline-concordant recommendations from bladder cancer multiparametric MRI (mpMRI) reports remains uncertain. Following TRIPOD-LLM, we retrospectively evaluated nine API-accessed LLMs using 590 mpMRI reports from three hospitals under two input conditions: original Chinese reports and clinician-translated, quality-controlled English reports. Two physicians independently rated each model-generated recommendation on a 1–5 ordinal scale. We analyzed 21,240 ratings using patient-level random-intercept cumulative link mixed models, with paired Wilcoxon tests as secondary exploratory analyses. The association between input condition and physician rating differed across LLMs (likelihood-ratio chi-square = 683.05, df = 8, P < 0.001). Compared with original Chinese input, clinician-translated English input was associated with higher ratings for chatgpt-o3-pro (OR 2.79, 95% CI 2.29–3.41), deepseek-v4-flash (OR 1.46, 95% CI 1.19–1.81), claude-opus-4-7 (OR 3.07, 95% CI 2.47–3.82), and gpt-5.5 (OR 1.88, 95% CI 1.53–2.32), but with lower ratings for claude-sonnet-4-5 (OR 0.29, 95% CI 0.23–0.36; all Holm-adjusted P < 0.001). Four LLMs showed no clear input-condition difference. Center-specific contrasts also depended on input condition and were materially unchanged after adjustment for age, sex, pathological T stage, tumor grade, tumor number, and maximum lesion diameter. Inter-rater agreement was substantial (quadratic weighted kappa range, 0.621–0.743). Readability did not track physician-rated recommendation quality. These findings argue against clinician-translated English as a universal optimization strategy and support model-specific validation of input conditions before clinical deployment. LLM-generated recommendations from mpMRI reports are best positioned as clinician-supervised, imaging-based decision support rather than standalone treatment advice.

npj Digital Medicine
Wuhan University (CN), Zhongnan Hospital of Wuhan University (CN), Second Hospital of Tianjin Medical University (CN), Tongji Hospital (CN), Hubei Cancer Hospital (CN), Huazhong University of Science and Technology (CN), Tianjin Medical University (CN)
National Natural Science Foundation of China, National Science and Technology Major Project
Good health and well-being
Artificial Intelligence in Healthcare and Education
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon