Effectiveness of Hybridsampling in Ensemble Learning for Thyroid Cancer Prediction

In recent years, Thyroid Carcinoma (TC) has become one of the most prevalent cancers. Diagnosing TC is difficult because its symptoms are very common with viral fever and there are overlap features between benign and malignant nodules. This encourages researchers to develop a Computer-Aided Diagnosis (CAD) system for TC diagnosis. However, a very common challenge in the CAD system for medical diagnosis is the problem of imbalanced datasets. This paper presents a comparative study on the effectiveness of different sampling techniques in handling the class imbalance problem for TC diagnosis. This study has explored six sampling techniques including oversampling, under-sampling, and hybrid-sampling, in combination with ten different machine learning models, including five single classifier-based models and five ensemble models. The experimental findings highlight the efficacy of hybrid sampling in addressing class imbalance for TC diagnosis where SMOTE-ENN in combination with Extra Trees classifier achieved accuracy and F1 scores of 92.2% and 89.7% respectively.