机器学习结合超声影像组学对上皮性卵巢癌BRCA基因突变亚型的预测价值  

The predictive value of machine learning combined with ultrasound imaging omics for BRCA gene mutation subtypes in epithelial ovarian cancer

作　　者：范懿 赵高芳 程瑞洪[1] FAN Yi;ZHAO Gaofang;CHENG Ruihong(Department of Medical Ultrasound,Sichuan Mianyang 404 Hospital,Mianyang,Sichuan 621000,China)

机构地区：[1]四川绵阳四0四医院医学超声科,四川绵阳621000

出　　处：《中国优生与遗传杂志》2024年第1期174-182,共9页Chinese Journal of Birth Health & Heredity

摘　　要：目的探讨机器学习结合超声影像组学对上皮性卵巢癌(EOC)患者乳腺癌易感基因(BRCA)突变亚型的预测价值。方法回顾性收集2022年2月至2023年6月收治于四川绵阳四0四医院行超声检查的105例EOC患者影像及病理资料,采用横断面分层法将患者分为建模组(70例)和验证组(35例)。将建模组患者按照BRCA基因是否突变分为突变组(22例)和野生组(48例),比较两组患者临床资料及影像学特征;并根据临床资料和影像组学特征建立临床(Clinical)模型、影像组学(Rad)模型和联合(Combine)模型对EOC患者BRCA基因突变进行预测。利用受试者操作特征(ROC)曲线、校准曲线和临床决策曲线对3种模型的预测价值进行评估。结果多因素Logistic回归分析模型显示,性别(OR=1.754,95%CI:1.573~1.942)、吸烟史(OR=1.611,95%CI:1.523~1.822)、家族史(OR=3.554,95%CI:1.324~5.684)为EOC患者BRCA基因突变的独立危险因素。根据LASSO回归方法对特征进行降尺度处理,并通过十倍交叉验证选择最优特征子集,共筛选出7个相关性较小的影像组学特征。建模组和验证组ROC曲线显示3种模型均具有良好的预测效能,Delong检验发现,建模组和验证组的Combine模型ROC曲线下面积均显著高于Rad和Clinical(P<0.05)。校准曲线显示各模型在建模组与验证组中均具有良好拟合效果。临床决策曲线显示Combine模型在建模组和验证组中较另两种模型具有较高的临床实用价值。结论机器学习和超声影像组学联合对EOC患者BRCA突变具有良好的预测价值。Objective To investigate the predictive value of machine learning combined with ultrasonography in breast cancer susceptibility gene(BRCA)mutations in patients with epithelial ovarian carcinoma(EOC).Methods The imaging and pathological data of 105 patients with EOC who underwent ultrasound examination in Sichuan Mianyang 404 Hospital from February 2022 to June 2023 were collected retrospectively.The patients were divided into model group(n=70)and verification group(n=35).The patients in the model group were divided into mutation group(n=22)and wild group(n=48)according to the mutation of BRCA gene.The clinical data and imaging features of the two groups were compared.According to clinical data and imaging characteristics,clinical(Clinical)model,imaging(Rad)model and combined(Combine)model were established to predict BRCA gene mutation in patients with EOC.The predictive value of the three models was evaluated by receiver operating characteristic(ROC)curve,calibration curve and clinical decision curve.Results Multivariate Logistic regression analysis showed that gender(OR=1.754,95%CI:1.573-1.942),smoking history(OR=1.611,95%CI:1.523-1.822)and family history(OR=3.554,95%CI:1.324-5.684)were independent risk factors for BRCA gene mutation in EOC patients.According to the LASSO regression method,the features were downscaling,and the optimal feature subset was selected by ten-fold cross-validation,and a total of 7 imaging features with less correlation were selected.The ROC curve of the modeling group and the verification group showed that the three models had good predictive efficiency.Delong test showed that the area under the ROC curve of the Combine model in the modeling group and verification group was significantly higher than that in Rad and Clinical(both P<0.05).The calibration curve shows that each model has a good fitting effect in the modeling group and verification group.The clinical decision curve shows that the Combine model has higher clinical practical value than the other two models in the modeling gro

关 键 词：BRCA基因 影像组学 上皮性卵巢癌 机器学习 

分 类 号：R737.31[医药卫生—肿瘤] R445.1[医药卫生—临床医学]