Addressing class imbalance with FSVM and interpreting key determinants in thyroid diagnosis through Explainable AI

Thyroid disease is recognized as one of the most prevalent medical diseases, disrupting the ability of the thyroid gland to function properly. Several machine learning algorithms like DT, KNN, LR, SVM, Random Forest, and XGBoost have been utilized to predict thyroid conditions. However, these models suffer from class imbalance and operate as black-box systems with restricted interpretability. This study aims to offer an efficient method named Fuzzy Support Vector Machine (FSVM) to address the class imbalance problem and reduce false positives in thyroid diagnosis. In addition, explainable AI approaches such as LIME and SHAP are used to increase the model interpretability and elucidate the predictive nature of the model. Linear and exponential FSVM have been applied to a thyroid dataset from Kaggle by assigning fuzzy memberships based on the distances from feature thresholds. The findings indicate that the FSVM enhances accuracy, with linear FSVM obtaining 96.65% and exponential FSVM achieving 97.22%, beating the standard SVM and other existing methods. Key clinical measurements such as TSH, FTI, and T4 are crucial for an accurate thyroid diagnosis. Moreover, pregnant women are at an increased risk of thyroid illness. Notably, the thyroxine dosage has been discovered as an essential treatment for thyroid patients.