Explainable machine learning predicts overall survival in female bladder cancer patients after radical cystectomy

Abstract

Background and purpose: Despite the higher incidence of bladder cancer in males, females face a disproportionately worse prognosis with more advanced disease. Accordingly, accurate survival prediction is crucial. This study sought to develop multiple machine learning models for predicting postoperative overall survival (OS) in female bladder cancer patients who have undergone radical cystectomy (RC).

Patients and methods: A retrospective analysis was conducted on female patients who underwent RC with postoperative pathological confirmation of bladder cancer in the SEER database from 2004 to 2022. These patients were randomly divided into a training set and an internal validation set at a 7:3 ratio. Additionally, 67 female bladder cancer patients from the Second Affiliated Hospital of Nanchang University were included as an external validation set. LASSO-Cox regression and Cox regression were used to identify independent prognostic factors for bladder cancer. Based on these factors, prediction models were constructed using five machine learning algorithms: Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Gradient Boosting Decision Tree (GBDT), Random Forest (RF), and XGBoost. The predictive performance of the models was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, precision, F1 score, concordance index (C-index), calibration curves, and decision curve analysis. The SHapley Additive exPlanations (SHAP) method was employed to interpret the most impactful features of the best-performing model.

Results: The study included a total of 4,603 patients. Lasso-Cox regression screening and Cox regression analysis revealed that T stage, N stage, age, tumor size, marital status, chemotherapy, and the number of examined lymph nodes (ELN) during surgery were significantly associated with the OS of female bladder cancer patients. Compared with other models, GBDT demonstrated superior discriminative ability in predicting 1-year, 3-year, and 5-year survival rates (1-year AUC(95%CI) = 0.771(0.742-0.801), 3-year AUC(95%CI) = 0.757(0.730-0.783), 5-year AUC(95%CI) = 0.745(0.721-0.770)), along with higher prediction accuracy, precision, C-index and F1 score. The calibration curve and decision curve analysis (DCA) confirmed the excellent predictive accuracy and clinical benefits of the GBDT model. These results were also validated in the external validation cohort. T stage, N stage, and chemotherapy were the most significant features of the optimal model, and SHAP analysis identified their important contributions within the model.

Conclusion: We have developed interpretable machine learning models to predict the OS of female bladder cancer patients following radical surgery. This model is intended to assist in clinical prognosis evaluation and provide a reference for individualized treatment decision-making.

Keywords: Female bladder cancer; Machine learning; Overall survival; Radical cystectomy; SEER.

Fig. 1

Research workflow diagram

Fig. 2

(A) Coefficient profiles of clinical and pathological features in the LASSO regression. (B) Feature selection using LASSO regression with 10-fold cross-validation based on the minimum criterion. The penalty parameter (λ) was applied to shrink coefficients toward zero for automatic feature selection. (C) LASSO regression coefficients of the key selected features

Fig. 3

ROC curves of SVM, KNN, GBDT, RF, and XGBoost models in the internal validation cohort (A, B, C) and external validation cohort (D, E, F)

Fig. 4

Calibration curves of SVM, KNN, GBDT, RF, and XGBoost models in the internal (A, B, C) and external (D, E, F) validation cohorts, and decision curve analysis (DCA) of these models in the internal (G, H, I) and external (J, K, L) validation cohorts

Fig. 5

Interpretability of the GBDT model evaluated using the SHAP method. (A) The SHAP bar plot displays feature importance based on mean absolute SHAP values. (B) The SHAP summary plot (beeswarm plot) illustrates the distribution of the impact of each feature on model predictions. Each point represents an individual patient, with colors indicating the feature value (red for high, blue for low). (C) The waterfall plot depicts the cumulative contribution path of features for a specific instance, starting from the base value to the model output