Development of a Multivariable Machine Learning Model for the Prediction of Postoperative Ileus After Radical Cystectomy.

BACKGROUND AND

OBJECTIVES

Postoperative Ileus (POI) is a common complication after Radical Cystectomy (RC) that delays recovery and extends hospital stay. Our objective was to develop machine learning (ML) models that can predict patients at a high risk of developing POI after RC.

METHODS

Data of patients who underwent RC for bladder cancer at our institution were retrospectively extracted for analysis and model creation. Data pre-processing and variable selection were applied. Several ML models were developed, evaluated, and compared using the area under the curve (AUC), F1 score, precision, sensitivity, and specificity.

RESULTS

Three hundred and seven patients were included in the analysis, and of these, 30 patients (9.8%) developed POI. Overall, 78.2% of patients were male and 91.2% were White. Of the various ML models, logistic regression demonstrated the highest AUC (0.98), followed by random forest (0.97), Support Vector Machine (0.97), XGBoost (0.95), simple neural network (0.87), and decision tree (0.84). Furthermore, the first two models also displayed higher specificity (0.94 and 0.93) and F1 scores (0.78 and 0.75). All models except for the decision tree and neural network achieved 100% sensitivity. Features such as age, body mass index (BMI), American Society of Anaesthesiologists (ASA) class status, surgical approach, potassium level, and the placement of a nasogastric tube were identified as key predictors of the outcome.

CONCLUSION

Based on our data, ML models can effectively predict POI after RC, especially with logistic regression and random forest. The identification of relevant predictors may contribute meaningfully to potential applications in risk stratification and personalized care.