Explainable machine learning to predict postoperative ileus after radical cystectomy: an 11-year real-world cohort

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Surgical Oncology

Background:

• Post-operative ileus (POI) is a common complication following radical cystectomy (RC).
• Existing risk assessment tools for POI have limited accuracy due to their inability to capture complex, non-linear relationships.
• Interpretable machine learning (ML) offers a promising avenue for enhancing early POI risk stratification.

Purpose of the Study:

• To develop and validate an interpretable ML model for early risk stratification of POI in RC patients.
• To compare the performance of various ML models in predicting POI.
• To identify key peri-operative factors contributing to POI risk.

Main Methods:

• A single-centre real-world cohort of 1,062 RC patients (2013-2023) was analyzed.
• Data included pre-operative comorbidities, medications, operative factors, and first-day laboratory indices.
• Five ML models were trained and validated, with feature importance assessed using SHAP (SHapley Additive exPlanations).

Main Results:

• POI occurred in 28.9% of patients.
• A back-propagation neural network demonstrated superior performance (AUC 0.828, accuracy 78.4%).
• Key predictors included intra-operative nasogastric tube placement, surgical approach, medication history, lymph-node dissection, lymphocyte count, and C-reactive protein.

Conclusions:

• An interpretable ML model utilizing routine peri-operative data accurately stratifies RC patients for POI risk from postoperative day 0.
• This ML approach outperforms traditional nomograms and highlights modifiable risk factors.
• Integration into EHR systems could facilitate real-time alerts and tailored patient management, pending multicentre validation.