BACKGROUND

Postoperative complications in colorectal surgery can significantly impact patient outcomes and healthcare costs. Accurate prediction of these complications enables targeted perioperative management, improving patient safety and optimizing resource allocation. This study evaluates the application of machine learning (ML) models, particularly deep learning neural networks (DLNN), in predicting postoperative complications following laparoscopic right hemicolectomy for colon cancer.

METHODS

Data were drawn from the CoDIG (ColonDx Italian Group) multicenter database, which includes information on patients undergoing laparoscopic right hemicolectomy with complete mesocolic excision (CME) and central vascular ligation (CVL). The dataset included demographic, clinical, and surgical factors as predictors. Models such as decision trees (DT), random forest (RF), and DLNN were trained, with DLNN evaluated using cross-validation metrics. To address class imbalance, the synthetic minority over-sampling technique (SMOTE) was employed. The primary outcome was the prediction of postoperative complications within 1 month of surgery.

RESULTS

The DLNN model outperformed other models, achieving an accuracy of 0.86, precision of 0.84, recall of 0.90, and an F1 score of 0.87. Relevant predictors identified included intraoperative minimal bleeding, blood transfusion, and adherence to the fast-track recovery protocol. The absence of intraoperative bleeding, intracorporeal anastomosis, and fast-track protocol adherence were associated with a reduced risk of complications.

CONCLUSION

The DLNN model demonstrated superior predictive performance for postoperative complications compared to other ML models. The findings highlight the potential of integrating ML models into clinical practice to identify high-risk patients and enable tailored perioperative care. Future research should focus on external validation and implementation of these models in diverse clinical settings to further optimize surgical outcomes.