Pre-Treatment Clinical and Biochemical Predictors of Therapy Response in Differentiated Thyroid Cancer Using Random Forest Models

The goal of this study was to construct a machine-learning framework to forecast patient response to radioiodine (131I) treatment and thyrotropin (TSH) suppression treatment in individuals with differentiated thyroid cancer (DTC) lacking structural evidence of disease, using only data available before treatment. In total, 597 patients were randomly selected for the training set to predict response to 131I therapy, while 326 were assigned for predicting response to TSH suppression therapy, all with DTC and no structural disease. Six different supervised machine-learning techniques were applied: Logistic Regression, Support Vector Machine, Random Forest (RF), Neural Networks, Adaptive Boosting, and Gradient Boost. These models were trained to identify effective response (ER) to 131I therapy and biochemical remission (BR) to TSH suppression therapy.  The key predictors of ER to 131I therapy were pre-treatment stimulated and suppressed thyroglobulin (Tg) values as well as radioiodine uptake before the ongoing 131I course. For Tg reduction during TSH suppression therapy, the main contributors were visible thyroid remnant on the post-treatment whole-body scan from the previous 131I course and TSH values. Random Forest (RF) outperformed the other algorithms. Using RF, the accuracy and area under the receiver operating characteristic curve (AUC) for differentiating ER from non-ER in 131I therapy reached 81.3% and 0.896, respectively. For forecasting BR during TSH suppression therapy, RF achieved an accuracy of 78.7% and an AUC of 0.857. These findings highlight the value of machine-learning approaches, particularly the Random Forest algorithm, as effective instruments for anticipating response to 131I therapy and TSH suppression therapy in DTC patients without structural disease, drawing solely on standard pre-treatment clinical parameters and laboratory indicators.