Machine learning predicts postoperative lung metastasis in upper tract urothelial carcinoma

Early detection of individuals at high risk for lung metastasis (LM) from upper tract urothelial carcinoma (UTUC) is important for postoperative risk stratification and surveillance planning. This retrospective study used five machine learning techniques to predict LM risk in UTUC patients. Data from the SEER database (2000–2017) were analyzed. The SEER cohort was randomly divided into a training set and an internal test set in a 7:3 ratio for model development and internal validation. No synthetic resampling technique was used in the final analysis. The Boruta algorithm and multivariable logistic regression identified key independent variables. The model’s performance was validated with data from 569 UTUC patients at the First Hospital of Jilin University. Evaluation included AUC, AUPRC, decision curve analysis, and SHAP interpretation. An online tool was developed for clinical use. Among the five machine learning models, RF showed strong apparent performance in some metrics, whereas LR demonstrated greater stability and generalizability across the internal test and external validation sets. Therefore, an LM-specific UTUC risk calculator was developed using LR for clinical application. This study developed a clinical prediction model for assessing LM risk in UTUC patients, offering potential support for clinical decision-making.