Development and validation of machine learning models for predicting muscle-invasive bladder cancer
Comparative performance analysis with VI-RADS
DOI:
https://doi.org/10.5489/cuaj.9728Keywords:
Bladder cancer, Muscle-invasive bladder cancer; machine learning, VI-RADS, Multiparametric mrı, Predictive modelingAbstract
Introduction: We aimed to develop and validate machine learning (ML)-based predictive models for the preoperative differentiation of muscle-invasive bladder cancer (MIBC) from non-muscle-invasive bladder cancer (NMIBC) using routinely available clinical and laboratory data, and to compare their diagnostic performance with the multiparametric magnetic resonance imaging (mpMRI)-based Vesical Imaging-Reporting and Data System (VI-RADS).
Methods: A total of 26 demographic, clinical, and biochemical variables were used to develop supervised ML models. Logistic regression, random forest, support vector machine, XGBoost, and LightGBM algorithms were trained to predict pathologic muscle invasion. The dataset was randomly divided into training (80%) and independent test (20%) cohorts. Model discrimination was evaluated using the area under the receiver operating characteristic curve (AUC). The performance of ML models was compared with a VI-RADS-based logistic regression model using the DeLong test, and model calibration was assessed using Brier scores.
Results: A total of 428 patients were included, comprising 317 with NMIBC and 111 with MIBC. In the independent test cohort, both the VI-RADS–based logistic regression model and the LightGBM model demonstrated high discriminative performance, with AUC values of approximately 0.93. Among ML methods, LightGBM showed the best overall performance. SMOTE improved sensitivity in most models, while calibration remained acceptable. DeLong analysis showed no significant difference between the LightGBM and VI-RADS models (p=0.49).
Conclusions: ML models based on routinely available clinical and laboratory variables achieved high diagnostic accuracy for predicting MIBC. The LightGBM model demonstrated performance comparable to the VI-RADS-based approach in the independent validation cohort and may serve as a complementary tool for risk stratification, particularly in settings where access to mpMRI is limited.
Downloads
Downloads
Published
How to Cite
Issue
Section
License
You, the Author(s), assign your copyright in and to the Article to the Canadian Urological Association. This means that you may not, without the prior written permission of the CUA:
- Post the Article on any Web site
- Translate or authorize a translation of the Article
- Copy or otherwise reproduce the Article, in any format, beyond what is permitted under Canadian copyright law, or authorize others to do so
- Copy or otherwise reproduce portions of the Article, including tables and figures, beyond what is permitted under Canadian copyright law, or authorize others to do so.
The CUA encourages use for non-commercial educational purposes and will not unreasonably deny any such permission request.
You retain your moral rights in and to the Article. This means that the CUA may not assert its copyright in such a way that would negatively reflect on your reputation or your right to be associated with the Article.
The CUA also requires you to warrant the following:
- That you are the Author(s) and sole owner(s), that the Article is original and unpublished and that you have not previously assigned copyright or granted a licence to any other third party;
- That all individuals who have made a substantive contribution to the article are acknowledged;
- That the Article does not infringe any proprietary right of any third party and that you have received the permissions necessary to include the work of others in the Article; and
- That the Article does not libel or violate the privacy rights of any third party.




