基于可解释性心脏磁共振参数的机器学习模型预测儿童心肌炎的预后

目的 探讨基于可解释性心脏磁共振(CMR)参数的机器学习模型对儿童心肌炎患者预后的预测价值。 方法 回顾性收集2012年9月至2017年11月临床诊断为儿童心肌炎患者45例,其中男28例,女17例,4~16岁,平均(9.8±3.4)岁。根据随访过程中是否出现心血管不良事件(ACE),将患者分为预后不良组(n=18例)和预后良好组(n=27例)。所有患者于住院治疗后进行CMR扫描,获取心功能、心肌应变、首过灌注及延迟强化(LGE)相关方面共206个可解释性的CMR参数。利用MATLAB分类学习应用程序对参数进行训练,挑选精度最高的模型作为预测模型。采用受试者工作特征曲线(ROC)对模型的预测效能进行评估。 结果 提取出14个可解释性的CMR参数,挑选其中无显著相关性的参数构建组合参数。单一参数中, 美国心脏协会(AHA)分段法中的第7节段最大信号强度百分比(SI %)预测性能最佳,曲线下面积(AUC)、预测敏感性和特异性分别为0.790、0.667和0.833;组合参数达到了最高的预测性能,AUC、预测敏感性和特异性分别为0.940、0.750和0.889。 结论 根据可解释性的CMR参数建立的机器学习模型对儿童心肌炎患者预后的预测具有良好价值,且在预后评估中组合参数比单一参数预测性能更高。

Machine learning models based on interpretive CMR parameters can predict the prognosis of pediatric myocarditis

Objective To develop and validate machine learning models based on interpretive cardiac magnetic resonance(CMR)parameters for prognosis evaluation of pediatric myocarditis. Methods A retrospective analysis of 45 pediatric patients with myocarditis was conducted. According to whether adverse cardiac events(ACE)occurred, the patients were divided into poor prognosis group(n=18)and good prognosis group(n=27). CMR scans were performed after hospitalization and 206 interpretive CMR parameters about myocardial function, myocardial strain, first-pass perfusion and late gadolinium enhancement(LGE)were obtained. The parameters were trained by the classification learner App in MATLAB and the training model with the highest accuracy was chosen as the best model. The receiver-operating characteristics(ROC)curve of the machine learning model was drawn to determine the prognostic performance. Results A total of 14 CMR parameters were selected as predictive factors, and those without correlation were used to construct the combined parameters. Among all these parameters, maximal signal intensity percentage(SI %)of the 7th segment of AHA had the best performance(AUC: 0.790, sensitivity: 0.667, specificity: 0.833). Combined parameters achieved the highest performance(AUC: 0.940, sensitivity: 0.750, specificity: 0.889). Conclusion The machine learning models based on interpretive CMR parameters can be used for prognosis evaluation of pediatric myocarditis, and combination of interpretive CMR parameters training with machine learning is more accurate than single ones.