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

Area of Science:

• Surgical Oncology
• Medical Informatics
• Predictive Analytics

Background:

• Accurate prediction of mortality risk following gastrectomy is crucial for optimizing surgical management and enhancing patient outcomes.
• Developing robust predictive models can aid in clinical decision-making for gastrectomy patients.

Purpose of the Study:

• To develop and compare machine learning (ML) models for predicting 30-day postoperative mortality after gastrectomy.
• To identify the key predictors of mortality in patients undergoing gastrectomy.

Main Methods:

• Utilized the NSQIP Participant Use Data File (2017-2022) to develop random forest, gradient-boosted tree, and XGBoost models.
• Compared ML models trained on comprehensive data (Model C) versus existing risk calculator variables (Model L).
• Evaluated model performance using the area under the receiver operating characteristics curve and identified predictors using SHapley Additive exPlanations.

Main Results:

• The XGBoost model demonstrated the highest predictive performance for 30-day mortality in both comprehensive (Model C) and limited (Model L) datasets.
• All developed ML models outperformed simple logistic regression in predicting mortality.
• Preoperative blood urea nitrogen and patient age were identified as the most significant predictors of 30-day mortality.

Conclusions:

• The XGBoost model offers superior predictive accuracy for 30-day postoperative mortality in gastrectomy patients.
• Preoperative laboratory values, specifically blood urea nitrogen, and age are critical factors influencing mortality risk.
• Integrating ML-based predictive models into clinical practice can enhance perioperative decision-making and improve outcomes for gastrectomy patients.