Predictive analytics for thyroid cancer recurrence: a feature selection and data balancing approach

Abstract Thyroid cancer recurrence presents considerable challenges in clinical practice, underscoring the need for accurate predictive models to guide timely interventions. This study introduces a hybrid machine learning (ML) framework that combines data balancing and feature selection to enhance recurrence prediction. Utilizing the Differentiated Thyroid Cancer Recurrence dataset, the framework evaluates the performance of nine distinct ML classifiers through an 80:20 stratified train-test split and stratified 5-fold cross-validation. Among the evaluated models, ensemble methods—particularly Random Forest and Bagging—demonstrate superior performance on SMOTE-balanced data, achieving $$98.7\%$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mrow> <mml:mn>98.7</mml:mn> <mml:mo>%</mml:mo> </mml:mrow> </mml:math> accuracy and $$95.5\%$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mrow> <mml:mn>95.5</mml:mn> <mml:mo>%</mml:mo> </mml:mrow> </mml:math> recall, and outperforming previously reported methods. Statistical analyses further confirm that the impact of feature selection techniques varies depending on classifier architecture. Overall, the proposed framework illustrates that combining data balancing with informed feature selection significantly enhances predictive performance and contributes to the development of reliable decision-support systems for the early detection of thyroid cancer recurrence. The framework’s interpretability and robustness underscore its potential for integration into clinical decision-support systems, enabling early recurrence detection and facilitating personalized treatment strategies. These findings are also applicable to other imbalanced medical datasets.