A radiomics-based interpretable model to predict post-cystectomy disease-free survival in muscle invasive bladder cancer: a multicenter retrospective cohort study

Abstract Objective To develop and validate an interpretable machine learning (ML) model based on CT radiomics to predict the 2-year disease-free survival (DFS) in muscle invasive bladder cancer (MIBC) patients with multicenter validation. Methods A total of 377 patients with MIBC who underwent radical cystectomy were included from five institutions between November 2012 and February 2022. Of these, 264 patients were assigned to the training set, and 113 to the test set. The least absolute shrinkage and selection operator (LASSO) regression was used to select radiomics features and construct radiomics score (Rad-score). Five common ML models were developed by combining the Rad-score with significant clinicopathologic factors. The shapley additive explanation (SHAP) method was employed to visualize the prediction process. Results The Rad-score was constructed by eleven radiomics features, achieving AUCs of 0.827 and 0.771 in predicting 2-year DFS in the training and test sets, respectively. Kaplan–Meier survival analysis showed significant differences in DFS between the high-risk and low-risk groups stratified by the Rad-score in both the training (hazard ratio: 5.40[95%CI: 3.89-7.49]) and test sets (hazard ratio: 6.11[95%CI: 3.39-11.02]). When combined with Rad-score, pT stage, variant histology, tumor size, and age, the RF model achieved the best performance, with AUCs of 0.924 and 0.850 in the training and test sets, respectively. SHAP analysis indicated that Rad-score and pT stage contributed the most to the RF model's predictions. Conclusions The proposed ML model demonstrated the potential to accurately predict early recurrence and mortality in MIBC patients, aiding personalized treatment decision-making.