Multimodal prognostic models for bladder urothelial carcinoma: uroplakin III combined with serum and demographic data

Runlin Feng ¹, Jian Hou ², Yanping Tao ³

⁰Department of Pathology, The Second Affiliated Hospital of Kunming Medical University, Kunming, Yunnan, China.

Frontiers in Oncology +

|

October 6, 2025

View abstract on PubMed

Summary

Integrating Uroplakin III (UPK3A protein) expression with inflammation and demographic data significantly improves bladder urothelial carcinoma (BUC) risk prediction using machine learning models for better adjuvant therapy guidance.

Area Of Science

Oncology
Biomarkers
Machine Learning

Background

Bladder urothelial carcinoma (BUC) is a recurrent and heterogeneous cancer.
Accurate postoperative risk stratification is essential for guiding adjuvant therapy.
Novel prognostic models are needed to improve patient management.

Purpose Of The Study

To develop advanced machine learning (ML) models for improved prognostic prediction in BUC.
To investigate the utility of Uroplakin III (UPK3A protein) expression combined with other factors.
To enhance postoperative risk stratification for BUC patients.

Main Methods

Retrospective analysis of 1,032 BUC patients undergoing radical cystectomy.
Collection of clinical, pathological, and serological data, including UPK3A protein expression.
Feature selection using LASSO regression and training/validation of nine ML models (e.g., LightGBM, RF, XGBoost).

Main Results

Machine learning models, particularly Random Forest (RF), showed high performance (AUCs: 0.894/0.754).
SHAP analysis identified vascular invasion, tumor necrosis, and UPK3A protein as key predictors.
Integrated models incorporating UPK3A protein outperformed traditional prognostic approaches.

Conclusions

Combining UPK3A protein expression with multimodal features significantly enhances BUC prognostic modeling.
This approach provides a promising clinical decision support tool for risk stratification and postoperative management.
Future research should explore transcriptomic/proteomic data for further validation.

Keywords:

BUC ML UPK3A protein multimodal data prognostic prediction risk stratification serum inflammatory markers shap