Interpretable machine learning model of contrast-enhanced CT radiomics for predicting post-BCG recurrence in...

While intravesical Bacillus Calmette-Guérin Vaccine (BCG) instillation remains standard adjuvant therapy for high-grade non-muscle-invasive bladder cancer (NMIBC) post-resection, marked interpatient response heterogeneity complicates recurrence prediction. This study develops a contrast-enhanced computed tomography (CT) radiomics-based machine learning model to quantify tumor heterogeneity and noninvasively predict 5-year recurrence in high-grade NMIBC.

This retrospective study included 136 patients with histopathologically confirmed high-grade NMIBC from our institution and an external cohort of 51 patients from an independent testing center. All patients underwent transurethral resection of bladder tumor (TURBT) followed by BCG instillation therapy. The internal cohort was randomly partitioned into a training set (n = 95) and a validation set (n = 41) at a 7:3 ratio, with an independent external test set (n = 51) used for external validation. All patients received contrast-enhanced CT prior to treatment. Independent clinicopathological predictors were identified through univariate and multivariate logistic regression analyses. Radiomics features were selected using the Least Absolute Shrinkage and Selection Operator (LASSO) regression, and a radiomics score (Radscore) was constructed. Four machine learning models-Gradient Boosting Decision Tree (GBDT), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), and Random Forest (RF)-were developed. Model performance was comprehensively evaluated via the area under the receiver operating characteristic curve (AUC), accuracy, precision, F1 score, confusion matrix, calibration curve, and decision curve analysis (DCA). Additionally, the interpretability of the models was assessed using Shapley Additive Explanations (SHAP).

Multivariate logistic regression analysis identified tumor size (OR = 3.11, 95% CI: 1.16-8.34, p = 0.024), number of lesions (OR = 4.56, 95% CI: 1.52-13.70, p = 0.007), and tumor calcification (OR = 3.41, 95% CI: 1.15-10.15, p = 0.027) as independent clinical predictors of 5-year recurrence following BCG instillation. A total of 4,738 radiomic features were extracted from contrast-enhanced CT images, and the top 20 most discriminative features were selected via LASSO regression to construct the Radscore. Among the four machine learning models developed by combining clinical factors and Radscore, SVM demonstrated the superior performance, with an AUC of 0.816 (95% CI: 0.774-0.858), accuracy of 73.2%, precision of 74.3%, and F1 score of 0.734. This performance was significantly better than that of XGBoost (AUC = 0.727, 95% CI: 0.683-0.770; accuracy 65.9%, precision 65.4%, F1 score 0.655), RF (AUC = 0.717, 95% CI: 0.675-0.752; accuracy 68.3%, precision 67.9%, F1 score 0.677), and GBDT (AUC = 0.685, 95% CI: 0.633-0.740; accuracy 70.7%, precision 71.4%, F1 score 0.709). External validation using independent test sets confirmed these.

SHAP analysis revealed that the number of lesions and Radscore were the most influential predictors in the SVM model. Calibration curves and DCA demonstrated that the SVM model exhibited robust stability and provided substantial clinical benefit.

We developed an interpretable machine learning model derived from contrast-enhanced CT radiomics, integrating clinicopathological characteristics and quantitative radiomic features to accurately predict the 5-year recurrence risk in patients with high-grade NMIBC following BCG instillation.