Development and validation of a machine learning model for predicting 30‐day major morbidity and mortality following radical cystectomy: An American College of Surgeons National Surgical Quality Improvement Program study

Abstract Objective To develop and validate machine learning models for predicting 30‐day major morbidity and mortality in patients undergoing radical cystectomy (RC) using the American College of Surgeons National Surgical Quality Improvement Project (ACS‐NSQIP) and to compare performance against a logistic regression model. Material and methods We identified 11 241 patients from the ACS‐NSQIP database 2020–2024 who underwent radical cystectomy. Demographics and comorbidities were extracted along with targeted variables from the NSQIP RC‐targeted database. The cohort was split into a training and an independent validation dataset. Predictive models were developed for logistic regression, random forest (RF) and XGBoost. Model performance was evaluated using the area under the receiver operating characteristic (ROC) curve, SHapley Additive exPlanations (SHAP) analysis, sensitivity, specificity and Brier scores. Results Of 11 241 patients, 2691 (23.9%) experienced at least one major complication, and 185 (1.6%) died within 30 days. Overall complications were 6365 (56.62%). Non‐survivors were significantly older (72.65 ± 10.14 vs. 68.50 ± 10.32 years, p < 0.001). Patients with major morbidity had lower mean albumin levels (3.92 ± 0.56 vs. 4.01 ± 0.48, p < 0.001). Logistic regression identified high BMI (OR 1.15, p < 0.001), black race (OR 1.34, p = 0.003), Hispanic ethnicity (OR 1.37, p = 0.009), prior pelvic surgery (OR 1.15, p = 0.002) and continent diversions (OR 1.46, p = 0.001) as predictors of 30‐day major morbidity, while low frailty (mFI‐5 ≤ 1; OR 0.72, p = 0.001) and higher preoperative albumin (OR 0.88, p < 0.001) were protective. For 30‐day mortality, increasing age (OR 1.42, p < 0.001) was the strongest risk factor. For 30‐day morbidity, the XGBoost model achieved the highest AUC 0.796 (95% CI: 0.783–0.814). For 30‐day mortality, the RF model showed superior discrimination with an AUC of 0.921 (95% CI: 0.908–0.934). SHAP analysis showed predictors of major morbidity were frailty, BMI and advanced age, whereas predictors of mortality were age, ASA class and preoperative creatinine levels. Decision curve analysis showed net clinical benefit for all three models. The web‐based tool can be accessed and used for prediction ( https://cystectomyai.streamlit.app/ ). Conclusions We developed and validated machine learning models for 30‐day major morbidity and 30‐day mortality following radical cystectomy. These findings support the integration of machine learning into clinical workflows to enhance preoperative counselling and personalized risk reduction.