Hybrid Feature Selection and Stacking Ensemble Learning with Dual Explainable AI to Esophageal Cancer Neoplasm Status Classification

Esophageal cancer is among the most deadly malignancies worldwide, with a survival rate of less than 25 percent, mostly due to late diagnosis. The correct and timely determination of neoplasm status is important for making therapy decisions and predicting patient outcomes. This paper will suggest a highly practical and explainable machine learning program to classify the neoplasm status (Tumor Free vs. with Tumor) of esophageal cancer. Using the Esophageal Cancer Clinical Dataset with 3,983 and 85 clinical features of patients with tumors in the study. Nine stacking ensemble methods, each combining eleven individual classifiers, namely: Decision Tree, XGBoost, Random Forest, Gradient Boosting, AdaBoost, MLP, SVM (LinearSVC), LightGBM, Logistic Regression, KNN and Naive Bayes, were systematically tested. A hybrid feature selection strategy based on unions with Mutual Information, 38 features of ANOVA F-test (SelectKBest), as well as 40 features of XGBoost importance, was used to reduce the 69 features to 40 features. The only aspect that was re-sampled was SMOTETomek on the training set to counter the imbalance in the classes. The optimal stacking model (Random Forest + Decision Tree, meta-learner: Logistic Regression) has the highest test accuracy (99.62%), F1-score (0.9962), and is able to make errors at a low rate (0.38). SHAP (beeswarm, bar, waterfall, dependence, and heatmap plots) and LIME (global aggregation, local instance-level, and four-sample grid analysis) were used two times, with clinical staging variables, survival indicators and lifestyle factors being the major predictors. Ablation experiment validated the performance improvements that are attributable to both the preprocessing elements. These results indicate the clinical promise of interpretable stacking ensemble learning to esophageal cancer decision support.