An ensemble learning model for predicting cancer-specific survival of muscle-invasive bladder cancer patients undergoing bladder preservation therapy

Abstract

Background: More muscle-invasive bladder cancer (MIBC) patients are now eligible for bladder-preserving therapy (BPT), underscoring the need for precision medicine. This study aimed to identify prognostic predictors and construct a predictive model among MIBC patients who undergo BPT.

Methods: Data relating to MIBC patients were obtained from the Surveillance, Epidemiology and End Results (SEER) database from 2004 to 2016. Eleven features were included to establish multiple models. The predictive effectiveness was assessed using receiver operating characteristic (ROC) curves, calibration plots, decision curve analysis (DCA) and clinical impact curve (CIC). SHapley Additive exPlanations (SHAP) were used to explain the impact of features on the predicted targets.

Results: The ROC showed that Catboost and Random Forest (RF) obtained better predictive discrimination in both 3- and 5-year models [test set area under curves (AUC) =0.80 and 0.83, respectively]. Furthermore, Catboost showed better performance in calibration plots, DCA and CIC. SHAP analysis indicated that age, M stage, tumor size, chemotherapy, T stage and gender were the most important features in the model for predicting the 3-year cancer-specific survival (CSS). In contrast, M stage, age, tumor size and gender as well as the N and T stages were the most important features for predicting the 5-year CSS.

Conclusions: The Catboost model exhibits the highest predictive performance and clinical utility, potentially aiding clinicians in making optimal individualized decisions for MIBC patients with BPT.

Keywords: Muscle-invasive bladder cancer (MIBC); bladder-preserving therapy (BPT); cancer-specific survival (CSS); ensemble learning; predictive model.

Figure 1

ROC curves of the four models: Catboost, Random Forest, Stacking, and Logistic Regression. Training sets for the 3- and 5-year CSS prediction models are shown in (A) and (B), respectively. Test sets for the 3- and 5-year CSS prediction models are shown in (C) and (D), respectively. ROC, receiver operating characteristic; CSS, cancer-specific survival.

Figure 2

Calibration plots for predicting 3- and 5-year CSS using different models. (A,B) Training sets of 3- and 5-year CSS predictive models, respectively. (C,D) Test sets of 3- and 5-year CSS predictive models, respectively. CSS, cancer-specific survival.

Figure 3

displays decision curve analysis graphs depicting the net benefit plotted against threshold probabilities. (A,B) Training sets of 3- and 5-year CSS predictive models, respectively. (C,D) Test sets of 3- and 5-year CSS predictive models, respectively. CSS, cancer-specific survival.

Figure 4

Clinical impact curves of the prediction model in the test set. (A) Three-year CSS models. (B) Five-year CSS models. The four solid curves indicated the number of cases predicted as deaths by the four models at each threshold probability; the four dashed lines refer to the actual number of deaths out of the number of cases predicted to die by the models at each threshold probability. CSS, cancer-specific survival.

Figure 5

The SHAP value of each feature was calculated based on the Catboost model. The bar charts depict the relative importance of the features. The beeswarm plots show the contribution of different features of all the patients to the predicted outcomes, with different colors representing the different values of a particular feature, and SHAP values greater than 0 representing a positive contribution to the predicted outcome. (A,B) Three-year CSS predictive model. (C,D) Five-year CSS predictive model. SHAP, SHapley Additive exPlanations; CSS, cancer-specific survival.