摘要
目的探讨影响人工智能心电分类算法抗扰性的因素。方法使用公开心电数据库和开源的人工智能心电分类算法模型,在算法训练、测试环节引入实验室测量的噪声,同时改变训练集的样本量,观察算法测试结果的变化趋势。结果当测试集单独添加噪声时,算法分类的总体准确性下降超过3%;当训练集、测试集分别添加相同类型噪声时,算法分类的总体准确性的降幅不超过0.5%;对于不同类型的心拍,训练集样本量对算法分类准确率的影响趋势各不相同。结论人工智能心电分类算法在训练阶段可引入必要的噪声,以加强算法本身的抗扰性,但同时应关注各分类之间的差异。训练集的扩增并非必然提升算法的抗扰性。
Objective To study factors that impacts immunity of artificial intelligence-enabled electrocardiograph classification algorithms.Methods By using the open electrocardiograph database and artificial intelligence-enabled electrocardiograph classification algorithm model,the noise measured in the laboratory was introduced into the algorithm training and testing,and the sample size of the training set was changed at the same time to observe the changing trend of the algorithm test results.Results When the test set was overlapped with noise,the overall accuracy of the algorithm classification was decreased more than 3%.When the same type of noise was added to the training set and the test set,the overall accuracy of the algorithm classification decreased by no more than 0.5%.For different types of heart beats,the sample size of training set had different influence trends on the classification accuracy of the algorithm.Conclusion Necessary noise can be introduced to the electrocardiograph classification algorithm in the training stage to enhance the algorithm’s immunity,but at the same time,attention should be paid to the differences between different classifications.The expansion of the training set does not necessarily improve the immunity of the algorithm.
作者
王浩
唐桥虹
唐娜
郝烨
李澍
孟祥峰
李佳戈
WANG Hao;TANG Qiaohong;TANG Na;HAO Ye;LI Shu;MENG Xiangfeng;LI Jiage(Institute for Medical Device Control,National Institutes for Food and Drug Control,Beijing 102629,China;College of Bioengineering,Chongqing University,Chongqing 400044,China)
出处
《中国医疗设备》
2023年第3期61-65,共5页
China Medical Devices
基金
国家重点研发计划(2019YFC0118802)。
关键词
人工智能医疗器械
鲁棒性
心电
神经网络
artificial intelligence medical device
robustness
electrocardiograph
neural network
