MACHINE LEARNING-BASED DOSE RECOMMENDATIONS FOR REMOVAL OF RESIDUAL THYROID TISSUE

Abstract

Thyroid cancer, particularly differentiated types such as papillary and follicular carcinomas, presents a significant health challenge globally and in Vietnam, where surgical removal of the thyroid gland (thyroidectomy) is the primary treatment approach. However, residual thyroid tissue often remains post-surgery, necessitating effective ablation to prevent recurrence and complications. Radioactive iodine therapy using I-131 is the standard method for ablating this residual tissue. The accuracy of I-131 dose estimation plays a crucial role in ensuring therapeutic success and patient safety. Traditional methods rely on general guidelines and physician expertise but may lack the precision necessary for individual patients. While several previous studies have attempted to improve dose estimation using machine learning approaches, they have often overlooked the importance of data preprocessing techniques. This lack of attention to data quality has limited the performance of predictive models. To address this problem, our study expands upon prior work by placing a strong emphasis on data preprocessing, aiming to enhance model accuracy and reliability. We implement a multi-stage framework that processes medical records and identifies key features, using seven traditional machine learning models for prediction. Our results demonstrate that the Decision Tree model outperforms other models, achieving the highest True Positive Rate of 0.995, a low False Positive Rate of 0.001, and exceptional performance in Recall, Precision, and F1-score (0.995, 0.996, and 0.995, respectively). Moreover, we developed decision rules generated by the Decision Tree for dose prediction, which use a dataset of clinical information from Vietnamese patients. This model represents a promising tool for improving radiation therapy delivery, ensuring more accurate, data-driven decisions, and ultimately better patient outcomes.