Incisional hernia prediction using machine learning models

One of the main complications after laparotomy is incisional hernia (IH), with an incidence of 40% in specific risk groups. There is no consensus on determining which patients are at low or high Risk. There is no currently full peer-reviewed study in which Machine learning (ML) is applied for the prediction of IH, so the objective of this study was to develop a predictive model of IH based on different ML techniques.

Retrospective cohort study type of diagnostic tests. Patients over 18 years of age who underwent midline laparotomy were included. The main outcome was the occurrence of IH. Three ML methods were evaluated: Logistic Regression, Decision Tree, and XGBoost. Each model's predictive capacity, Friedman range score, and clinical utility were assessed. The usefulness of all models was further evaluated using Bayes' theorem to determine the change in prevalence after applying the models.

789 patients were analyzed, 161 (20.1%) presented IH, the model was built with 13 predictor variables, identifying preoperative risk of surgical site infection as the strongest risk factor. For the XGBoost model, the best performance with an AUC of 0.93 ± 0.02 and a Brier score of 0.10 was achieved. Its clinical utility, confirmed by Bayes' theorem, showed a positive test raises the posterior probability to 86% (above treatment threshold), while a negative test reduces it to 2%.

A predictive model was created using the XGBoost technique. The robustness of the model is related to the correct selection of the variables. Cross-validation and a learning curve indicated that the model performs robustly and shows potential for generalization to new data. An easy-to-use web application was developed that does not require sophisticated or expensive clinical studies.

The study was registered with the hospital research and research ethics committees of the Regional Hospital of High Specialty of Bajio IMSS Bienestar under registration number CEI/HRAEB/002/2021. It was also registered on the Clinical Trials platform with registration number NCT05718999 on February 27^th, 2023.