Prediction model for postoperative complications in gastrointestinal surgery based on preoperative and intraoperative factors using machine learning: a retrospective, single-center study.

PURPOSE

Machine learning (ML) is a method of creating models by learning latent patterns and features from collected data to predict and classify new unknown data. We constructed an ML model to predict postoperative complications using various pre- and intraoperative factors from electronic medical records and examined its prediction accuracy.

METHODS

Data of 617 patients who underwent major organ resection were included in this study. Patient information was collected from the medical records. Consequently, we created Dataset 1, which included all of the data, and Dataset 2, which was adjusted for the factors. The ML models were applied to the two datasets, and the performances of the ML models were compared.

RESULTS

In Dataset 1, the Logistic Regression model showed the best performance, with correct predictions of serious postoperative complications (accuracy), area under the receiver operating characteristic curve (AUROC), and area under the precision-recall curve (AUPRC) values of 0.798, 0.671, and 0.374, respectively. The random forest (RF) model performed best in Dataset 2, with accuracy, AUROC, and AUPRC values of 0.855, 0.725, and 0.412, respectively.

CONCLUSIONS

The RF model performed well in predicting serious postoperative complications of gastrointestinal surgery. Further studies are required to improve the accuracy of this ML model for clinical applications.