Noninvasive prediction of metastasis in esophageal cancer using ensemble-based feature selection

Esophageal cancer (EC) is a significant health concern worldwide, and predicting its metastatic progression is essential for planning effective treatments. Histopathological intervention is the gold standard for diagnosing Esophageal Cancer Metastasis (ECM). However, we introduced a noninvasive, data-driven approach utilizing different machine learning (ML) algorithms on clinical data from TCGA to predict the risk of ECM. Among these algorithms, CatBoost stands out, achieving a 73% accuracy and a 75% area under the curve (AUC) using 5-fold cross-validation with a standard deviation of 4% among 5-folds. We visualized feature importance graphs and feature correlations to explain the decision-making of ML models. Our findings highlight associations between height, weight, age, alcohol consumption, the number of packs smoked, tumor location, and the risk of metastasis in EC patients. This approach offers a promising way to enhance EC metastasis prediction while minimizing invasive procedures.