Exploring the Prognostic Efficacy of Machine Learning Models in Predicting Adenocarcinoma of the Esophagogastric Junction

Abstract

Objective: To investigate the value of machine learning and traditional Cox regression models in predicting postoperative survivorship in patients with adenocarcinoma of the esophagogastric junction (AEG). Methods: This study analyzed clinicopathological data from 203 patients. The Cox proportional risk model and four machine learning models were constructed and internally validated. ROC curves, calibration curves, and clinical decision curves (DCA) were generated. Model performance was assessed using the area under the curve (AUC), while calibration curves determined the fit and clinical significance of the model. Results: The AUC values of the 3-year survival in the validation set for the Cox regression model, extreme gradient boosting, random forest, support vector machine, and multilayer perceptron were 0.870, 0.901, 0.791, 0.832, and 0.725, respectively. The AUC values of 5-year survival in the validation set for each model were 0.915, 0.916, 0.758, 0.905, and 0.737, respectively. The internal validation AUC values for the four machine learning models, extreme gradient boosting, random forest, support vector machine, and multilayer perceptron, were 0.818, 0.772, 0.804, and 0.745, respectively. Conclusion: Compared with Cox regression models, machine learning models do not need to satisfy the assumption of equal proportionality or linear regression models, can include more influencing variables, and have good prediction performance for 3-year and 5-year survival rates of AEG patients, among which, XGBoost models are the most stable and have significantly better prediction performance than other machine learning methods and are practical and reliable.