An Ensemble Learning System Based on Stacking Strategy for Survival Risk Prediction of Patients with Esophageal Cancer

Background: Predicting the prognosis of esophageal cancer (EC) patients is crucial for optimizing the treatment plan and allocating medical resources effectively.

Methods: This study proposes a novel ensemble learning-based EC survival prediction model. Firstly, recursive feature elimination (RFE) is used to determine the key feature subsets from the dataset. Based on the determined key features, the improved clustering by fast search and find of density peaks (IDPC) is proposed to construct a novel indicator related to EC survival risk. The cosine distance is introduced in IDPC to cluster samples with similar characteristics. Then, the adaptive synthetic (ADASYN) technique is used to generate more high-risk samples to balance high-risk and low-risk samples. Finally, the hyperparameters of the three models, including extreme gradient boosting (XGBoost), adaptive boosting (AdaBoost), and random forest (RF), are optimized by whale optimization algorithm (WOA) and a new stacking model is constructed to evaluate the survival risk of patients.

Results: The proposed stacking model achieved an area under the receiver operating characteristic curve (AUC) of 0.952 and Accuracy of 0.899, on the dataset from the First Affiliated Hospital of Zhengzhou University.

Conclusions: The survival prediction model the proposed ensemble learning system is much more accurate and convenient, providing a basis clinical judgment and decision making and improving the survival status of esophageal cancer patients.