An explainable machine learning model for predicting bladder tumor aecurrence risk

TL;DR

This study developed an accurate and transparent machine learning model to predict the risk of bladder tumor recurrence after surgery, using patient features like BMI and tumor size.

Contribution

The novel contribution is an explainable XGBoost model with high accuracy for bladder tumor recurrence prediction, validated in a clinical cohort.

Findings

XGBoost with seven features achieved an AUC of 0.994 in predicting bladder tumor recurrence.

BMI and maximum tumor diameter were the most influential features in the model.

The model provides transparent feature contributions, aiding clinical decision-making.

Abstract

Bladder cancer is associated with considerable postoperative recurrence rates. Accurate risk prediction remains challenging in clinical practice.

To develop and validate an explainable machine learning model for predicting bladder tumor recurrence following surgical treatment.

This retrospective cohort study enrolled 504 patients with pathologically confirmed bladder tumors treated at the Department of Urology, Zhejiang Dinghai Hospital, from October 2018 to October 2024. Postoperative surveillance was conducted at 3, 6, 12, and 24 months to assess recurrence status. The dataset was randomly partitioned into training (n=352) and testing (n=152) sets prior to analysis. LASSO regression with lambda.1se criterion was performed exclusively on the training set to identify predictive features, yielding 19 candidate variables. Subsequently, eleven machine learning algorithms were evaluated: Logistic Regression, Random Forest, XGBoost, Gradient Boosting Machine, Neural Network, AdaBoost, Decision Tree, C5.0, Support Vector Machine, Elastic Net, and Naive Bayes. Model performance was assessed using area under the receiver operating characteristic curve (AUC), recall, accuracy, F1-score, precision, and negative predictive value (NPV), with 95% confidence intervals calculated for all metrics.

During follow-up, 90 of 504 patients (17.9%) developed tumor recurrence. XGBoost utilizing seven features demonstrated optimal performance, achieving an AUC of 0.994 in the independent testing set. The final predictive variables included BMI, maximum tumor diameter, tumor morphology, smoking status, extravesical invasion signs, tumor number, and dome location. SHAP analysis identified BMI (mean absolute SHAP value: 1.5359) and maximum tumor diameter (1.4565) as primary contributors to predictions, followed by morphology (1.3370) and smoking status (1.2798).

The seven-feature XGBoost model provides accurate prediction of bladder tumor recurrence with transparent feature contributions. This explainable approach may assist clinicians in risk stratification and individualized surveillance planning.

Figures