采用多阅片者多病例设计评估AI辅助医疗产品临床试验的样本量估算和应用  

Estimation and Application of Sample Size in Clinical Trials of AI Assisted Medical Products by Multi-reader Multi-case Design

作　　者：尚美霞 阎小妍[2] 李雪迎[1] 朱赛楠[1] 姚晨[1,2] Shang Meixia;Yan Xiaoyan;Li Xueying(Department of Medical statistics,Peking University First Hospital(100034),Beijing;不详)

机构地区：[1]北京大学第一医院医学统计室,100034 [2]北京大学临床研究所

出　　处：《中国卫生统计》2022年第1期14-18,共5页Chinese Journal of Health Statistics

基　　金：国家重点研发计划主动健康和老龄化科技应对重点专项(2020YFC2006400)。

摘　　要：目的探索基于多阅片者多病例研究设计的人工智能辅助医疗器械或软件进行确证性临床试验的病例和阅片者样本量的选择依据和方法。方法以某三甲医院收集的某胸部CT影像辅助诊断系统用于肺结节诊断的试验数据为例,采用R软件对数据进行方差分析和诊断准确度统计,以方差和协方差分析的结果为依据对计划开展的正式试验样本量进行估算,同时探索分析不同参数组合条件下的样本量变化情况。结果7名具有不同胸部CT阅片经验的影像科医生对回顾性纳入的200例病例影像分别在AI辅助条件和无AI辅助条件下进行读片,采用AI辅助医生读片的诊断准确性优于医生独立读片的准确性[AF-ROC差值0.119(95%CI:0.078,0.161),P<0.001],不同阅片者提高的准确度水平不同。同一阅片者分别在两种诊断模式下诊断结果的相关系数越大,需要病例样本量越小;相同阅片模式下不同阅片者间诊断结果的相关系数和不同阅片者在不同诊断模式间诊断结果的相关系数的差值越大,需要的病例样本量越大。结论基于多阅片者多病例研究设计的人工智能或机器学习辅助医疗类的器械或软件的确证性临床试验中采用多阅片者会适当减低病例的样本量,采用合适的统计学方法可以对阅片者和病例的样本量进行科学的估算。Objective To explore the method of case and reader sample size for confirmatory clinical trials with AI assisted medical devices or software based on multi-reader multi-case study design.Methods Taking the test data of a chest CT image aided diagnosis system used in the diagnosis of pulmonary nodules as an example,the R software was used to analyze the variance and diagnostic accuracy of the data.Based on the results of the analysis of variance and covariance,the sample size of the planned formal test was estimated,and the change of sample size under different parameter combinations was explored and analyzed.Results Seven radiologists with different chest CT film reading experience read 200 cases under AI assisted condition and without AI assisted condition respectively.The diagnostic accuracy of AI assisted doctor reading was better than that of doctor independent reading(AFROC-AUC difference 0.119(95%CI:0.078,0.161),P<0.001).In terms of the relationship between the correlation coefficient and the sample size,the larger the correlation coefficient of the diagnosis results of the same reader in the two diagnostic modes,the smaller the sample size of the cases needed;the larger the difference between the correlation coefficient of the diagnosis results of different readers in the same mode and the correlation coefficient of the diagnosis results of different readers in different diagnostic modes,the larger the sample size of the cases needed.Conclusion In the confirmatory clinical trials of AI or machine learning aided medical devices or software based on multi reader and multi case study design,using multi readers will reduce the sample size of cases appropriately,and using appropriate statistical methods can scientifically estimate the sample size of readers and cases.

关 键 词：多阅片者多病例 人工智能辅助 病例 样本量 

分 类 号：R195.1[医药卫生—卫生统计学]