A Statistically Robust Machine Learning Framework for Thyroid Dysfunction Prediction Using Classical Classifiers

Abstract—Thyroid dysfunction is a common endocrine problem whose symptoms may overlap with many non-thyroid conditions, making timely screening important. This paper presents a conference-style classical machine-learning framework for predicting thyroid dysfunction from structured clinical and laboratory variables. Logistic Regression, Support Vector Machine, Random Forest, Decision Tree, k-Nearest Neighbors, and Gaussian Naive Bayes are compared under a leakage-aware preprocessing and validation protocol. The framework emphasizes rigorous statistical evaluation through stratified nested cross-validation, bootstrap confidence intervals, receiver-operating-characteristic analysis, probability calibration, DeLong comparison of AUCs, and McNemar testing of paired classification errors. Accuracy, Precision, Recall, Specificity, and F1-score are reported with their mathematical definitions. In the illustrative project experiment, Random Forest provides the strongest overall discrimination and balanced error profile, while Logistic Regression remains a transparent baseline and calibrated SVM is competitive. Feature analysis identifies thyroid-stimulating hormone and free-thyroxine-related variables as the dominant predictors. A lightweight web prototype demonstrates patient/laboratory input and real-time prediction output for hypothyroid-pattern and euthyroid cases. The framework is intended for research and screening support rather than autonomous clinical diagnosis.