基于机器学习构建放射性碘治疗疗效的预测模型

Construction of predictive models of radioiodine therapy based on machine learning

目的基于分化型甲状腺癌(DTC)患者的临床资料及放射学参数,通过机器学习算法构建放射性碘治疗(RAI)疗效的预测模型。方法选取2015年12月至2020年12月于山东大学第二医院核医学科接受RAI治疗的1642例DTC患者为研究对象,在RAI治疗结束的6个月后评估其治疗疗效,筛选与疗效相关的核心特征进行机器学习建模。将研究对象按就诊时间(2019年7月)划分为训练集(n=973)和验证集(n=669),于训练集中利用Logistic、随机森林、支持向量机、Adaboost 4种方法进行模型构建,利用受试者工作特征曲线下面积(AUC)、灵敏度和特异度评估模型的性能,绘制校准曲线及决策曲线评估模型的准确度和临床受益性,并在验证集中评估模型外部稳定性。结果4种模型的预测性能较高,稳定性较好,预测精度和净收益高于目前临床常规应用的肿瘤-淋巴结-转移(TNM)分期和复发风险分层。Logistic模型表现最佳,其AUC在训练集中为0.827、验证集中为0.869。结论基于机器学习构建的术后RAI治疗疗效预测模型有较高的预测性能,由此构建的列线图可实现个体化精准预测。

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.