Deep learning algorithm using fundus photographs for 10-year risk assessment of ischemic cardiovascular diseases in China  

基于眼底照片估算缺血性心血管病10年风险的深度学习算法

作　　者：Yanjun Ma Jianhao Xiong Yidan Zhu Zongyuan Ge Rong Hua Meng Fu Chenglong Li Bin Wang Li Dong Xin Zhao Jili Chen Ce Rong Chao He Yuzhong Chen Zhaohui Wang Wenbin Wei Wuxiang Xie Yangfeng Wu 马燕军;熊健皓;朱一丹;戈宗元;花蓉;付萌;李承龙;王斌;董力;赵昕;陈吉利;荣策;和超;陈羽中;汪朝晖;魏文斌;解武祥;武阳丰(Peking University Clinical Research Institute,Peking University First Hospital,Beijing 100191,China;PUCRI Heart and Vascular Health Research Center at Peking University Shougang Hospital,Beijing 100191,China;Key Laboratory of Molecular Cardiovascular Sciences(Peking University),Ministry of Education,Beijing 100191,China;Beijing Airdoc Technology Co.,Ltd.,Beijing 100081,China;Beijing Tongren Eye Center,Beijing Tongren Hospital,Beijing 100005,China;Shibei Hospital,Shanghai 200435,China;iKang Guobin Healthcare Group Co.,Ltd.,Beijing 100022,China)

机构地区：[1]Peking University Clinical Research Institute,Peking University First Hospital,Beijing 100191,China [2]PUCRI Heart and Vascular Health Research Center at Peking University Shougang Hospital,Beijing 100191,China [3]Key Laboratory of Molecular Cardiovascular Sciences(Peking University),Ministry of Education,Beijing 100191,China [4]Beijing Airdoc Technology Co.,Ltd.,Beijing 100081,China [5]Beijing Tongren Eye Center,Beijing Tongren Hospital,Beijing 100005,China [6]Shibei Hospital,Shanghai 200435,China [7]iKang Guobin Healthcare Group Co.,Ltd.,Beijing 100022,China

出　　处：《Science Bulletin》2022年第1期17-20,M0003,共5页科学通报（英文版）

基　　金：supported by the National Natural Science Foundation of China(81974489 and 81974490)。

摘　　要：缺血性心脑血管病(ischemic cardiovascular diseases,ICVD)包括缺血性卒中与缺血性心脏病.传统的基于危险因素的ICVD风险预测模型在实践推广中受限,本文旨在开发并验证一种利用眼底照片估算10年ICVD风险的深度学习算法,用于替代传统模型.算法的开发和验证均以传统模型所预测的10年ICVD风险为参照.研究基于体检人群(390,947人)数据开发卷积神经网络算法.在内部验证(20,571人)中,该算法估算10年ICVD风险的自然对数的调整R^(2)为0.876,筛查临界/中等及以上(≥5%/≥7.5%)ICVD风险人群的受试者工作特征曲线(ROC)下面积(AUC)分别为0.971(95%CI:0.967~0.975)和0.976(95%CI:0.973~0.980);在中老年人群的外部验证(1309人)中,调整R^(2)为0.638,AUC分别为0.859(95%CI:0.822~0.895)和0.876(95%CI:0.816~0.937).该算法有望替代传统ICVD风险预测模型,用于在基层医疗机构进行ICVD风险快速筛查,但仍需前瞻性研究进行验证.

关 键 词：风险预测模型 基层医疗机构 缺血性卒中 缺血性心脑血管病 缺血性心脏病 体检人群 中老年人群 卷积神经网络 

分 类 号：R54[医药卫生—心血管疾病]