Predicting Ileus after Colorectal Surgery Using Machine Learning

Introduction: Prolonged postoperative ileus (PPOI) is one of the most common complications after colorectal operation. Predicting PPOI is helpful for guiding postoperative decision-making. We sought to develop predictive models using machine learning, an approach that can identify non-linear patterns in large datasets. Methods: We used colectomy and proctectomy cases from the American College of Surgeons National Surgical Quality Improvement Program database. The primary outcome was PPOI. Various machine learning techniques were applied, including random forest (RF), gradient boosting (XGB), and neural networks (NN), and compared with logistic regression (LR). Each model was evaluated using area under the receiver operating characteristic curve (AUROC) and area under the precision recall curve (AUPRC). Results: The dataset included 276,232 patients with 46,774 (17%) experiencing PPOI. All machine learning models performed similarly, with NN and XGB both showing AUROC 0.766 (95% CI 0.760-0.772) and RF showing AUROC 0.763 (95% CI 0.757-0.769); this is significantly higher than LR which achieved AUROC 0.717 (95% CI 0.71-0.723). For AUPRC, machine learning models also performed similarly, with NN scoring 0.431 (95% CI 0.424-0.437), XGB scoring 0.43 (95% CI 0.424-0.436), and RF scoring 0.427 (95% CI 0.421-0.433). In contrast, LR scored 0.359 (95% CI 0.353-0.365). The features identified by NN as most important for model decision-making were sex, operative time, operative approach, age, procedure type, and ostomy placement. Conclusion: Machine learning techniques significantly outperformed LR in the prediction of PPOI. With external validation, these models could be used to identify patients at highest risk of PPOI and guide postoperative decision-making.