A predictive model for L-T4 dose in postoperative DTC after RAI therapy and its clinical validation in two institutions.

Purpose: To develop a predictive model using machine learning for levothyroxine (L-T4) dose selection in patients with differentiated thyroid cancer (DTC) after resection and radioactive iodine (RAI) therapy and to prospectively validate the accuracy of the model in two institutions. Methods: A total of 266 DTC patients who received RAI therapy after thyroidectomy and achieved target thyroid stimulating hormone (TSH) level were included in this retrospective study. Sixteen clinical and biochemical characteristics that could potentially influence the L-T4 dose were collected; Significant features correlated with L-T4 dose were selected using machine learning random forest method, and a total of eight regression models were established to assess their performance in prediction of L-T4 dose after RAI therapy; The optimal model was validated through a two-center prospective study (n=263). Results: Six significant clinical and biochemical features were selected, including body surface area (BSA), weight, hemoglobin (HB), height, body mass index (BMI), and age. Cross-validation showed that the support vector regression (SVR) model was with the highest accuracy (53.4%) for prediction of L-T4 dose among the established eight models. In the two-center prospective validation study, a total of 263 patients were included. The TSH targeting rate based on constructed SVR model were dramatically higher than that based on empirical administration (Rate 1 (first rate): 52.09% (137/263) vs 10.53% (28/266); Rate 2 (cumulative rate): 85.55% (225/263) vs 53.38% (142/266)). Furthermore, the model significantly shortens the time (days) to achieve target TSH level (62.61 ± 58.78 vs 115.50 ± 71.40). Conclusions: The constructed SVR model can effectively predict the L-T4 dose for postoperative DTC after RAI therapy, thus shortening the time to achieve TSH target level and improving the quality of life for DTC patients.