Fuglsang Anders
Hiort Lorenzens Vej 6c, DK6100 Haderslev, Denmark.
Eur J Pharm Sci. 2021 Jan 1;156:105595. doi: 10.1016/j.ejps.2020.105595. Epub 2020 Oct 11.
In recent years regulators have documented how pharmaceutical companies or clinical research organisation can manipulate bioequivalence trial data for non-approvable formulations by performing an interim analysis followed by re-analysis of pharmacokinetic profiles under new subject aliases, with a switch of Test and Reference and/or dilutions. The net effect is that point estimates for failing products will be forced artifically towards 1 and that trials will pass the test for bioequivalence. This is not detectable by any pharmacopoeial method, and is not addressed by common assessment practices at agencies. This paper aims at demonstrating how the signals of such fraudulent study conduct can be detected. The approaches presented are called "Buster" and "SaToWIB" routines; these are computer programs that have been used extensively by regulators to detect signals of fraud but they have not been described in the public domain. The Buster routines visualize trends in the form of partial statistics, residual plots, cumulative confidence intervals, cumulative mean squared errors, and more. Runs tests on the sign of the residuals may constitute a potential test for the manipulation. It is noteworthy that in 2020, regulators in the European Union have publicly begun questioning trial validity on basis of PK profile similarity. The SaToWIB routines rank profile pairs according to numerical similarity on basis of an objective function. It is shown that the rank (as determined by score) is an indicator of fraud in that the actual fraud cases will have higher rank than if there were no relationship between rank and score. The paper also comments on the use of multivariate statistics and discusses the need for development of formal tests for manipulation in view of e.g. multiplicity.
近年来，监管机构已记录了制药公司或临床研究组织如何通过进行中期分析，然后以新的受试者别名重新分析药代动力学特征，并切换试验药物和参比药物及/或稀释度，来操纵不可批准制剂的生物等效性试验数据。最终结果是，失败产品的点估计值将被人为地推向1，并且试验将通过生物等效性测试。任何药典方法都无法检测到这一点，而且各机构的常规评估做法也未涉及这一问题。本文旨在说明如何检测此类欺诈性研究行为的信号。所介绍的方法称为“克星”（Buster）和“萨托维布”（SaToWIB）程序；这些是监管机构广泛用于检测欺诈信号的计算机程序，但尚未在公共领域进行描述。“克星”程序以部分统计量、残差图、累积置信区间、累积均方误差等形式直观呈现趋势。对残差符号进行游程检验可能构成对操纵行为的潜在检验。值得注意的是，2020年，欧盟监管机构已公开开始基于药代动力学特征相似性对试验有效性提出质疑。“萨托维布”程序根据目标函数基于数值相似性对特征对进行排序。结果表明，排序（由分数确定）是欺诈的一个指标，因为实际的欺诈案例将比排序与分数之间没有关系时具有更高的排名。本文还对多元统计的应用进行了评论，并鉴于例如多重性等问题，讨论了开发用于检测操纵行为的正式检验方法的必要性。
