Incisional hernia prediction using machine learning models.

Area of Science:

• Surgical Oncology
• Medical Informatics
• Data Science

Background:

• Incisional hernia (IH) is a common complication after laparotomy, affecting up to 40% of patients in high-risk groups.
• Current methods lack consensus for identifying patients at high risk for IH.
• Machine learning (ML) has not been extensively studied for IH prediction.

Purpose of the Study:

• To develop and evaluate ML models for predicting IH after midline laparotomy.
• To identify key risk factors for IH development.
• To assess the clinical utility of predictive models using Bayes' theorem.

Main Methods:

• Retrospective cohort study including 789 adult patients undergoing midline laparotomy.
• Evaluation of three ML techniques: Logistic Regression, Decision Tree, and XGBoost.
• Assessment of model performance using AUC, Brier score, and clinical utility via Bayes' theorem.

Main Results:

• 161 out of 789 patients (20.1%) developed IH.
• The XGBoost model achieved an AUC of 0.93 ± 0.02 and a Brier score of 0.10.
• Preoperative surgical site infection risk was the strongest predictor; the model accurately reclassified risk probabilities.

Conclusions:

• An XGBoost-based ML model effectively predicts IH risk post-laparotomy.
• The model demonstrates robustness and potential for generalization, validated by cross-validation and learning curves.
• An accessible web application was developed for practical clinical use.