Machine learning models to predict pathological complete response after chemoradiotherapy in muscle-invasive bladder cancer: toward optimized indications for bladder preservation

After bladder-sparing chemoradiotherapy (CRT) for muscle-invasive bladder cancer (MIBC), residual microscopic disease may cause intravesical recurrence. This study aimed to develop machine learning models to predict pathological complete response (pCR) after CRT.

We retrospectively analyzed 285 patients with cT2-4N0M0 MIBC who underwent CRT followed by partial or radical cystectomy. CRT response was determined from cystectomy specimens. Patients were randomly assigned to a training cohort (n = 228) and a test cohort (n = 57) using a fixed random seed. A total of 102 machine learning models were evaluated in the training cohort. The model with the highest area under the receiver operating characteristic curve (AUC) in the training-cohort was selected and then evaluated in the test cohort. The model incorporated clinical T stage, tumor size, multiplicity, hydronephrosis, bladder neck involvement (BNI), concomitant carcinoma in situ (CIS), variant histology, grade, and progression.

Clinical T stages were T2/3/4 in 169/105/11 patients, respectively. Hydronephrosis was observed in 69 patients. Transurethral resection before CRT revealed multiplicity, BNI, and concomitant CIS in 108, 54, and 34 patients, respectively. Overall, 178/285 patients (62%) achieved pCR. WeightedEnsemble_L4 was selected as the final model and had a test-cohort AUC of 0.829 (95% confidence interval [CI], 0.706-0.932). The calibration slope was 0.856 (95% CI, 0.469-1.639), and the wide CI indicated uncertainty in calibration.

Machine learning based on clinicopathologic variables estimated pCR after CRT. This may support decision-making regarding bladder preservation in patients with MIBC.