Evaluation of Classical and Ensemble Machine Learning Algorithms for Thyroid Cancer Diagnosis: A Comparative Evaluation

Thyroid cancer is a growing global health concern, necessitating reliable and accurate diagnostic tools to support early detection and clinical decision-making. This study aims to develop and implement classical and ensemble machine learning models based on clinical, demographic, and biochemical data to predict thyroid cancer risk. Pearson correlation analysis was employed to identify and select the most relevant features for model training. A range of classifiers was optimized using hyperparameter tuning and cross-validation strategies. To assess robustness and generalizability, model performance was evaluated using accuracy, precision, recall, and F1-score across two independent datasets. Results show that ensemble models, particularly CatBoost, Bagging (Random Forest), and XGBoost, achieved the highest performance, with accuracies of up to 98.70% and F1-scores of 0.99 on Dataset 2, while maintaining consistent performance on Dataset 1 with accuracies around 82.51%. Classical models such as Logistic Regression, LDA, and SVM also performed competitively, achieving up to 97.40% accuracy on Dataset 2 and 82.51% on Dataset 1. These findings demonstrate the effectiveness of combining feature selection with optimized machine learning models and highlight the potential of ensemble approaches for improving thyroid cancer risk as- sessment in clinical practice.