关键词: 放射性碘治疗, 机器学习, 分化型甲状腺癌, 预测模型, 治疗疗效

Abstract: Objective To construct the predictive models for radioactive iodine therapy(RAI)efficacy by machine learning algorithms based on clinical data and radiological parameters of patients with differentiated thyroid cancer(DTC). Methods A total of 1,642 DTC patients treated with RAI during Dec. 2015 and Dec. 2020 were collected. The efficacy was evaluated 6 months after RAI and the core features associated with efficacy were screened out for machine-learning modeling. The patients were divided into the training set(n=973)and validation set(n=669)according to the consultation time(July 2019). In the training set, the Logistic model, random forest model, support vector machine model and Adaboost model were constructed, whose performances were evaluated with the area under the receiver operating characteristic curve(AUC), accuracy and specificity. The calibration and decision curves were drawn to evaluate the accuracy and clinical benefits of the models. The external stability was assessed in the validation set. Results The four predictive models had great predictive performance, having higher predictive accuracy and net benefits than the tumor-node-metastasis(TNM)staging and recurrence risk stratification. The Logistic model was the most effective, with an AUC of 0.827 in the training set and 0.869 in the validation set. Conclusion Predictive models for RAI efficacy built on machine learning have great predictive performance. The developed nomogram can realize individualized and accurate prediction for the therapeutic efficacy of RAI for DTC.

Key words: Radioactive iodine therapy, Machine learning, Differentiated thyroid cancer, Predictive model, Therapeutic efficacy