Development of a machine learning-based predictive model for transitional cell carcinoma of the renal pelvis in White Americans: a SEER-based study.

Background: Transitional cell carcinoma (TCC) of the renal pelvis is a rare cancer within the urinary system. However, the prognosis is not entirely satisfactory. This study aims to develop a clinical model for predicting cancer-specific survival (CSS) at 1-, 3-, and 5-year for White Americans with renal pelvic TCC.

Methods: Data of all White American patients diagnosed with TCC of the renal pelvis from 2010 to 2015 were extracted and analyzed from the Surveillance, Epidemiology, and End Results (SEER) database in this retrospective study. Subsequently, after excluding the metastatic group, a subgroup analysis was performed on the data of 1,715 White Americans with non-metastatic renal pelvic TCC. Patients included in this study were randomly divided into the training and validation sets in a ratio of 7:3. In addition, the features in the training set were extracted by the Boruta algorithm. The importance of these features was visualized using the eXtreme Gradient Boosting (XGBoost)-based SHapley Additive exPlanation (SHAP) tool. To improve predictive accuracy, a nomogram model with these identified independent prognostic variables was developed.

Results: A total of 1,887 White American patients with renal pelvic TCC were included in this study. In the training set, the area under the curve (AUC) for CSS nomograms at 1-, 3-, and 5-year were 0.813 [95% confidence interval (CI): 0.774-0.852], 0.738 (95% CI: 0.702-0.774), and 0.733 (95% CI: 0.698-0.768), respectively. Correspondingly, the AUCs for CSS nomograms at the above time points were 0.781 (95% CI: 0.732-0.830), 0.785 (95% CI: 0.741-0.829), and 0.775 (95% CI: 0.729-0.820) in the validation set, respectively. The subgroup analysis results revealed that the AUCs for CSS nomograms at 1-, 3-, and 5-year were 0.788, 0.725, and 0.726 in the training set, respectively, while the AUCs for CSS nomograms at 1-, 3-, and 5-year were 0.831, 0.786, and 0.754 in the training set, respectively.

Conclusions: In this study, a nomogram that predicts CSS in White American patients diagnosed with renal pelvic TCC was efficiently constructed. The application of the nomogram may enhance patient care and assist clinicians in choosing the optimal treatment strategies.