Bladder cancer AI preprint shows why clinical risk models need direction checks
Published in Artificial Intelligence, Journals, Tools.
Published October 5, 2026. Source date: October 1, 2026.
An arXiv preprint describes a prediction model that assigned lower bladder cancer recurrence risk when tumour stage or carcinoma in situ worsened. Conventional measures of discrimination and calibration did not reveal this clinically implausible behaviour.
The analysis used the UK BOXIT trial: 472 patients enrolled across 51 sites, with 435 having at least two years of follow-up. Researchers changed one clinical feature at a time to test the direction of the predicted risk. In the unconstrained model, the direction reversed in 90.2% of carcinoma-in-situ comparisons and 74.3% of tumour-stage comparisons.
Adding prespecified monotonic constraints removed those violations while maintaining similar discrimination: AUC 0.723 versus 0.718. The result supports checking whether a model responds sensibly to clinically meaningful changes, alongside standard performance metrics.
This preprint uses one trial and internal-external validation. Independent prospective validation remains necessary before clinical deployment.