Feature Importance-Guided Ensemble Classification for Predicting Recurrence in Differentiated Thyroid Cancer

Abstract

Accurate prediction of cancer recurrence is critical for improving patient monitoring and personalized treatment planning. In this study, we propose a machine learning framework to predict recurrence in patients with differentiated thyroid cancer using statistically selected clinical features. Feature relevance was assessed using ANOVA for ordinal/numerical variables and the Chi-square test for one-hot encoded categorical variables, allowing us to identify the most informative predictors. We then trained three distinct classifiers—Random Forest, Logistic Regression, and XGBoost—and combined them using a hard voting ensemble strategy. The proposed ensemble achieved an accuracy of 98.7% on the test set, with particularly strong precision and recall scores for the recurrent class, indicating its potential clinical utility. Interestingly, all three base classifiers produced identical predictions on the test data, suggesting the dataset’s strong internal structure and the effectiveness of our feature selection process. This work highlights the value of integrating statistical feature selection with ensemble modeling for robust and interpretable prediction in clinical oncology applications.