基于多参数MRI影像组学及临床特征的鼻咽癌远处转移可解释性机器学习预测模型

Prediction of distant metastasis in nasopharyngeal carcinoma by interpretable machine learning model based on multiparametric MRI radiomics and clinical factors

目的 建立基于多参数MRI影像组学及临床特征的机器学习预测模型,并评价其治疗前预测鼻咽癌(nasopharyngeal carcinoma, NPC)远处转移风险的效能及临床应用价值。材料与方法 回顾性分析2010年6月至2017年9月来自三家医院的1393例经病理证实的NPC患者的临床资料及MRI图像(训练队列1049例、外部验证队列344例)。用ITK-SNAP勾画感兴趣区并用Pyradiomics包逐层提取特征。使用相关性分析、单因素分析和递归特征消除法筛选特征,最后通过梯度提升机(Gradient Boosting Machine, GBM)算法构建模型。通过受试者工作特征(receiver operating characteristic, ROC)曲线比较模型的预测效能,以及决策曲线分析评估临床实用性。利用SHAP(SHapley Additive exPlanation)算法赋予最佳预测模型可解释性。结果 经筛选后最终保留10个影像组学特征。基于影像组学特征、临床特征、影像组学+临床特征三种特征组合构建了GBM_R、GBM_C和GBM_RC模型。三者在训练集上的ROC曲线下面积(area under the curve, AUC)值分别为0.938、0.724和0.938;GBM_RC(命名为NPC-Wise)在外部验证集中取得了最高的AUC值,为0.775。N分期是NPC-Wise预测过程中最重要的特征。SHAP模型预测力图能直观可视化特征影响NPC-Wise预测远处转移风险的过程。结论 基于多参数MRI影像组学及临床特征的可解释性机器学习预测模型NPC-Wise在预测NPC远处转移风险方面具有较高效能,SHAP算法为其提供了个体水平的可解释性,能为个性化治疗提供有价值的决策依据。

Objective: To establish a machine learning prediction model based on multi-parametric MRI features and clinical variables, and evaluate its efficacy in predicting distant metastasis in nasopharyngeal carcinoma(NPC) before treatment. Materials and Methods: MRI images of 1393 patients with pathologically confirmed NPC from three hospitals(1049 in the training cohort and 344 in the external validation cohort) from June 2010 to September 2017 were retrospectively analyzed. We used ITK-SNAP and Pyradiomics to delineate regions of interest and extract radiomic features, respectively. Features were selected using correlation analysis, univariate analysis and recursive feature elimination(RFE) method. The gradient boosting machine(GBM) algorithm was utilized to construct models. Receiver operating characteristic(ROC) curve and area under the curve(AUC) were used to compare the predictive efficacy of the models, and decision curve analysis(DCA) was used to assess the clinical utility. The SHapley Additive exPlanation(SHAP) algorithm was used to attribute interpretability to the optimal prediction model. Results: Ten radiomic features were finally selected. GBM_R, GBM_C and GBM_RC models were constructed based on the three features combination: radiomic features, clinical variables, and radiomic features + clinical variables. The AUC values of the them on the training set were 0.938, 0.724, and 0.938, respectively. GBM_RC(hereafter, NPC-Wise)achieved the highest AUC value of 0.775 in the external validation set. The SHAP force plot provided a visualization of the direction and degree of influence of each feature on the predicting results of the model. Conclusions: The interpretable machine learning prediction model NPC-Wise, based on multi-parametric MRI radiomic features and clinical variables, showed good performance in predicting the risk of distant metastasis in NPC, as well as providing individual-level interpretability with the SHAP algorithm, which can provide a valuable decision basis for personalized treatment.