Deciphering Oncological Recurrence through Hybridized Machine Learning Paradigms in Thyroid Disorders

Abstract

Thyroid disorders are manifesting with increasing frequency, rendering early identification indispensable for mitigating mortality and associated complications. Precise estimation of disease progression and elucidation of interrelationships among clinical attributes are paramount for efficacious medical diagnosis and therapeutic interventions. This investigation tackles these imperatives by leveraging machine learning and deep learning paradigms, augmented by sophisticated clinical feature analysis and ensemble-learning methodologies. Such an integrative framework facilitates both the prognostication of disease trajectory and the evaluation of interdependencies among clinical determinants. We deployed a quintet of models - K-Nearest Neighbors (KNN), XGBoost, Random Forest, Convolutional Neural Network (CNN), and TabNet - to discern the optimal prognostic outcomes for thyroid conditions. Empirical findings reveal that the proposed methodologies decisively eclipse conventional approaches in predictive performance. The CNN model attained the pinnacle of accuracy (98.70%) and Area Under the Curve (AUC) metrics (100%), trailed by TabNet, which yielded 97.40% accuracy and 99.18% AUC, underscoring their adeptness in addressing intricate classification challenges. Furthermore, an ensemble machine learning classifier amalgamating multiple models realized 100% sensitivity alongside 99.71% accuracy. This ensemble approach, adept at surmounting feature dimensionality reduction and pronounced class imbalance, substantiates its efficacy in predicting thyroid anomalies with unparalleled precision.