Machine Learning Model Predicts Bladder Cancer Recurrence After Surgery

Researchers may have found a more precise way to predict whether bladder cancer will return after surgery. A multicenter study published in Scientific Reports found that machine learning models drawing on patient characteristics, tumor pathology, lab results, and postoperative follow-up data could help identify patients at higher risk for recurrence.

Postoperative recurrence is a major determinant of prognosis in bladder cancer. Even after standard treatment with transurethral resection of bladder tumor (TURBT), recurrence is common, but identifying high-risk patients early could help clinicians tailor follow-up schedules and treatment plans more precisely after surgery.

Current risk-assessment tools rely heavily on tumor features such as stage, size, grade, and number. While those factors remain important, they may not fully capture individual recurrence risk, particularly when postoperative biomarkers and inflammatory signals are not included.

A Personalized Approach

The retrospective study included 488 adult patients with non–muscle-invasive bladder cancer who were treated at Wuxi People’s Hospital or Wuxi Second People’s Hospital in China between January 2018 and January 2023. All patients had histologically confirmed disease and underwent primary TURBT. Researchers excluded patients with prior bladder surgery, previous systemic therapy, other primary malignancies, severe comorbidities, active urinary tract infection, or incomplete follow-up data.

Recurrence was defined as histologically confirmed urothelial carcinoma detected in the bladder during the 2-year follow-up period, excluding residual tumors found within 3 months of surgery and cases that progressed to muscle-invasive disease.

Investigators collected 32 variables from each patient for a comprehensive view of postoperative recurrence risk, including demographic characteristics, lifestyle factors, comorbidities, tumor characteristics, clinical features, surgical details, postoperative treatment, and laboratory markers measured 3 months after surgery. These markers included neutrophil-to-lymphocyte ratio (NLR), urine cytology, hematuria, NMP22, and alkaline phosphatase.

The Search for the Best Model

The research team developed and tested four AI models based on machine learning algorithms: eXtreme Gradient Boosting (XGBoost), Random Forest, Support Vector Machine, and k-Nearest Neighbors. To assess performance they focused on discrimination, calibration, and clinical utility by looking at receiver operating characteristic curves, calibration curves, decision curve analysis, and 10-fold cross-validation. The researchers also used SHapley Additive exPlanations (SHAP) analysis, a method for showing how much each variable contributed to a model’s prediction, to clarify which factors were driving the results.

The best-performing model were also tested on an external validation data set to determine the generalizability and applicability of the model on an independent cohort from Tengzhou Central People’s Hospital.

Several factors were independently associated with recurrence by univariate and multivariate analysis. Patients aged 65 years or older had higher odds of recurrence, as did patients with a history of smoking. Tumor-related predictors included T1 stage, multifocal disease, tumor size of at least 3 cm, and high-grade pathology. Postoperative markers also had predictive value, including an NLR of at least 3, positive urine cytology, hematuria, positive NMP22, and alkaline phosphatase of at least 120 U/L.

XGBoost performed best among the four models. It achieved an area under the receiver operating characteristic curve (AUC) of 0.960 in the training set and 0.925 in the validation set. In the validation set, the model had an accuracy of 0.850, sensitivity of 0.825, specificity of 0.865, and F1 performance score of 0.807. In the external validation cohort, XGBoost had an AUC of 0.850, suggesting that the model retained predictive performance when applied to an independent patient group.

Understanding the Predictions

The SHAP analysis showed that the most influential features included urine cytology, pathologic grade, NMP22, hematuria, smoking history, tumor number, tumor stage, tumor size, and NLR.

Rather than relying on a single risk factor, the model appeared to build a broader risk profile by combining traditional tumor characteristics with postoperative clinical and laboratory signals. “The inclusion of hematological and urinary biomarkers not only enhances the model’s biological interpretability but also provides new insights into the mechanisms underlying bladder cancer recurrence,” the study authors, including co-corresponding author Chunyang Chen, PhD, Department of Urology, The Affiliated Wuxi People’s Hospital of Nanjing Medical University, Wuxi People’s Hospital, Wuxi Medical Center, Nanjing Medical University, Wuxi, China, wrote.

The findings suggest that machine learning could eventually support more personalized surveillance after TURBT by helping clinicians identify patients who may need closer follow-up. Still, the authors noted several limitations: the study was retrospective, and laboratory markers may be affected by infection, medication use, or timing of measurement. The model was also not directly compared with established clinical prediction tools such as the European Organization for Research and Treatment of Cancer risk tables or the Club Urológico Español de Tratamiento Oncológico scoring model.

For clinicians, the practical question is whether postoperative biomarkers belong in routine risk assessment alongside standard pathologic features. According to the investigators, this model suggests they might, but larger prospective studies are needed for confirmation.

DISCLOSURES: This study was supported by “Taihu Light” Science and Technology Innovation Project (Basic Research) of Wuxi, Wuxi Municipal Health Commission Youth Science Fund, General projects of Wuxi Medical Center, Nanjing Medical University and Nanjing Medical University Science and Technology Development Fund-General Project, Major Scientific Research Projects of Wuxi City. The study authors reported no conflicts of interest.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.