Machine Learning to Predict Early Death Despite Pancreaticoduodenectomy

Area of Science:

• Oncology
• Machine Learning
• Surgical Outcomes

Background:

• Pancreaticoduodenectomy (PD) for right-sided pancreatic ductal adenocarcinoma (PDAC) has a 25% 1-year mortality rate.
• A significant portion of patients experience morbidity without survival benefits from PD compared to non-surgical options.

Purpose of the Study:

• To compare the accuracy of machine learning models against traditional regression models in predicting futile surgery for PDAC patients.
• To identify key preoperative factors associated with futile PD.

Main Methods:

• Analysis of National Cancer Database data (2004-2020) for PDAC patients undergoing PD.
• Definition of futile PD as death within 12 months of cancer diagnosis.
• Training and testing of logistic regression, multilayer perceptron, decision tree, random forest, and gradient boosting models using 16 preoperative variables.

Main Results:

• Out of 66,331 patients, 25.3% met criteria for futile surgery.
• Gradient boosting model achieved the highest accuracy (AUC 0.689), outperforming logistic regression, random forest, and decision tree.
• Predictors of futile PD included advanced age, larger tumor size, and poor differentiation; neoadjuvant therapy reduced futility risk.

Conclusions:

• Machine learning models demonstrate moderate accuracy in predicting futile PD for PDAC patients.
• Findings support improved shared decision-making and optimized care strategies for PDAC.
• Further research with more granular data is warranted.