Advancing postoperative mortality prediction in gastrectomy: a machine learning approach using NSQIP data

Accurate prediction of mortality risk in gastrectomy is critical to optimize surgical management and improve patient outcomes. This study aims to develop machine learning (ML) models for predicting 30-day postoperative mortality following gastrectomy and identify key predictors.

We utilized the National Surgical Quality Improvement Program (NSQIP) Participant Use Data File from 2017 to 2022 to develop ML models: (1) Random Forest, (2) Gradient-Boosted Tree, and (3) XGBoost model. A simple logistic regression model was further developed to compare the model's prediction. We trained each model using a comprehensive set of variables available in the NSQIP data (Model C) or 17 variables included in the existing American College of Surgeons (ACS) NSQIP risk calculator (Model L). We used the area under the receiver operating characteristic curve to evaluate the model performance and employed SHapley Additive exPlanations algorithms on the best-performing model to identify the most impactful predictors.

Of 7954 patients who underwent gastrectomy, approximately 4.3% of patients died within 30 days following gastrectomy. In both Model C and Model L, the XGBoost model showed the best performance, followed by the Random Forest. The Model C outperformed the Model L, and all ML models outperformed simple logistic regression. In the XGBoost model, preoperative blood urea nitrogen was the most important predictor, followed by age.

The XGBoost model demonstrated the highest predictive performance for 30-day postoperative mortality following gastrectomy. Preoperative laboratory variables and age were key predictors. Incorporating ML-based models into clinical practice has the potential to enhance perioperative decision-making and improve patient outcomes after gastrectomy.