Machine Learning for Predicting the Low Risk of Postoperative Pancreatic Fistula After Pancreaticoduodenectomy: Toward a Dynamic and Personalized Postoperative Management Strategy.

Area of Science:

• Surgical Oncology
• Medical Informatics
• Biostatistics

Background:

• Postoperative pancreatic fistula (POPF) is a significant complication after pancreaticoduodenectomy (PD).
• Existing predictive models for POPF lack accuracy and fail to integrate early postoperative data.
• This study addresses the need for improved POPF prediction using advanced computational methods.

Purpose of the Study:

• To develop and validate machine learning (ML) models for predicting the absence and severity of POPF.
• To identify key clinical, surgical, and early postoperative variables influencing POPF risk.
• To enhance the accuracy of POPF prediction beyond conventional approaches.

Main Methods:

• Retrospective analysis of 216 patients undergoing PD.
• Systematic evaluation of 24 ML algorithms, including GradientBoostingClassifier.
• Application of SHAP analysis for model interpretability and identification of critical predictors.

Main Results:

• GradientBoostingClassifier demonstrated superior predictive performance compared to traditional logistic regression.
• The ML model effectively identified patients at low risk for POPF in the early postoperative period.
• Early postoperative biomarkers were identified as key predictors in the ML model.

Conclusions:

• ML models show promise in accurately stratifying POPF risk after PD.
• These models can potentially optimize postoperative management, including early drain removal.
• External validation is crucial to confirm generalizability and clinical utility for personalized surgical care.