Xu Zhiheng, Li Meijuan
1 U.S. Food and Drug Administration, Silver Spring, MD, USA.
Ther Innov Regul Sci. 2019 Sep;53(5):623-629. doi: 10.1177/2168479018804175. Epub 2018 Oct 31.
The gold standard in conducting clinical trials/studies is to follow what is prespecified in the study protocol. However, deviations from the study protocol may occur. This article discusses the issues of protocol deviation in pivotal clinical trials or studies for medical device and provides statistical approaches to mitigating bias such as selection bias specifically for diagnostic test clinical trials or studies.
Bias correction methods are developed for 2 specific types of selection biases, prescreening bias and verification bias. Statistical approaches are discussed on how to estimate device performance adjusted for enrollment enrichment and discrepant testing results. We use an FDA-approved Roche Cobas Human Papillomavirus (HPV) test for detecting high-grade cervical disease (>CIN2) as an example to illustrate how to correct for verification bias. A recently FDA-cleared Microarray Assay in detecting copy number variation is used to illustrate how to properly estimate sensitivity and specificity for the discrepancy analysis.
The unadjusted sensitivity and specificity based on verified samples were 83.2% and 60.4% for the Roche's HPV test. However, using the correction method with the missing-at-random assumption, the verification bias-adjusted sensitivity and specificity were 34.5% and 93.6%, respectively.
Protocol deviations can lead to biased estimates of device clinical performance if not handled appropriately. Statistical methods correcting for bias and protocol deviations are recommended in estimating device performance.
开展临床试验/研究的金标准是遵循研究方案中预先规定的内容。然而，可能会出现与研究方案的偏差。本文讨论了医疗器械关键临床试验或研究中的方案偏差问题，并提供了减轻偏差的统计方法，如针对诊断测试临床试验或研究的选择偏倚。
针对两种特定类型的选择偏倚，即预筛选偏倚和验证偏倚，开发了偏差校正方法。讨论了如何估计针对入组富集和不一致检测结果进行调整后的器械性能的统计方法。我们以美国食品药品监督管理局（FDA）批准的罗氏 Cobas 人乳头瘤病毒（HPV）检测高级别宫颈疾病（>CIN2）为例，说明如何校正验证偏倚。以最近 FDA 批准的用于检测拷贝数变异的微阵列分析为例，说明如何正确估计差异分析的敏感性和特异性。
罗氏 HPV 检测基于验证样本的未调整敏感性和特异性分别为 83.2%和 60.4%。然而，使用基于随机缺失假设的校正方法，校正验证偏倚后的敏感性和特异性分别为 34.5%和 93.6%。
如果处理不当，方案偏差可能导致器械临床性能的估计有偏差。在估计器械性能时，建议采用校正偏差和方案偏差的统计方法。
