Machine Learning Classifiers for Radioactive Iodine Therapy Decision-Making in Thyroid Cancer

Introduction: Radioactive iodine therapy (RAIT) is crucial for treating patients with differentiated thyroid carcinoma after the initial surgery. However, determining if RAIT is necessary and the appropriate iodine dose can be challenging, especially for inexperienced physicians. This study developed and compared machine learning classifiers to aid RAIT decision-making for thyroid cancer patients.
Methods: The study included a cohort of 210 patients who had undergone total thyroidectomy. Two machine learning classifiers were trained to suggest if RAIT was necessary and to propose an appropriate I-131 dose. The classifiers were evaluated using accuracy and the kappa coefficient to measure agreement with the gold standard decision made by an experienced physician. To test the classifiers’ acceptance by potential users, groups of 30 patients were presented to four young nuclear medicine physicians, who were asked to propose the best therapy treatment for each patient before and after using the classifier as a decision support system.
Results: The Artificial Neural Network (ANN) algorithm demonstrated higher accuracy (95.71%) and kappa coefficient (0.96, range: 0.91-1.00) than the Naïve Bayes (NB) classifier. The kappa coefficient increased from 0.70 to 0.93 when four young nuclear medicine physicians used the ANN classifier as a decision support tool, indicating its usefulness for educational purposes.
Conclusion: The machine learning classifiers developed in this study can aid inexperienced medical professionals in decision-making on RAIT for thyroid cancer patients. The ANN algorithm outperformed the NB classifier and can be a reliable tool for determining if RAIT is necessary and proposing an appropriate I-131 dose for thyroid cancer patients after the initial surgery.