Privacy-Preserving Federated Learning for Thyroid Disease Detection: A Comparative Study on Feature Selection and Class Balancing

Detecting thyroid disorders early and correctly is very important for good medical care. Real clinical data from the UCI repository is used in this study to show a full federated learning system for identifying thyroid diseases. The technique includes a lot of data preprocessing, like handling categorical and numerical features, missing values, and normalisation. It also includes five cutting-edge class balancing methods (SMOTE, BorderlineSMOTE, ADASYN, RandomOverSampler, and SMOTEENN) and two advanced feature selection techniques (SelectKBest using ANOVA F-score and Recursive Feature Elimination with logistic regression). Each balanced and feature-selected datasetwas given to five federated clients so that they could work together to train a deep neural network using the Flower federated learning framework. This kept patient data private. The model’s performance was measured over 10 communication rounds using F1-score, accuracy, precision, recall, and the combination of SelectKBest and SMOTEENN. This combination achieved the highest F1-score (0.9123) and accuracy (0.9133) in finding cases of the minority class. These results showhowimportant it is for shared healthcare AI to include feature selection and class balance. SHAP was used to explain model predictions; T3, TSH, and FTI emerged as the strongest contributors, consistent with clinical knowledge. A permutation-importance analysis based on AUROC drop produced the same ordering, providing an independent, model-agnostic check. This will allow for a scalable and private way to accurately screen for thyroid problems in clinical settings.