Modelling of ROC curves in meta-analysis of diagnostic test accuracy studies and network meta-analysis

本项目涉及医学证据合成的两个活跃研究领域，诊断测试准确性研究的meta分析和网络meta分析。最终的目标是将这两个领域结合起来。在诊断准确性研究的荟萃分析的标准方法中，假设每个研究贡献一对敏感性和特异性。在项目的第一阶段，我们研究了两种更一般的方法，它们解释了潜在生物标志物的多个阈值，“建模生物标志物分布”和“平均ROC曲线”。第一种方法是基于在非患病和患病个体中估计潜在生物标志物的分布函数的想法。基于分布假设，我们利用线性混合效应模型对适当转换的数据估计两组的分布参数。该模型既考虑了研究内敏感性和特异性的依赖性，也考虑了研究间的异质性。我们得到了一个总的接受者工作特征（SROC）曲线，以及在每个特定阈值下的综合灵敏度和特异性。此外，通过最大化约登指数，确定跨研究的最佳阈值是可能的。第二种方法最初由Martinez-Camblor（2014）提出，通过在“垂直”方向上平均（即平均灵敏度，以特异性为条件），将汇总ROC曲线确定为对主要研究的ROC曲线的加权平均。我们通过(1)交换敏感性和特异性的作用，即以敏感性为条件对特异性进行平均（水平平均），以及(2)以其和为条件对真阳性率和假阳性率（即约登指数）的差异进行平均（对角线平均）来扩展该方法。另一个研究领域是网络元分析（NMA）。在四篇论文中，我们研究了(1)NMA与电网络理论之间的关系，(2)通过适当扩大标准误差来调整多臂研究的替代方法，(3)网络自动可视化方法，以及(4)基于网络元分析的频率主义者治疗排名。在第二个项目阶段，我们希望完善和扩展所有这些方法，特别关注知识翻译。一方面，先进的方法越来越多，另一方面，它们对非统计用户的可及性受到限制。我们希望通过编写一个新的R包来实现“建模生物标志物分布”和“平均ROC曲线”方法，并扩展我们现有的R包netmeta来进行网络元分析，例如通过元回归，从而弥合这一差距并继续遵循这一目标。最终目的是将DTA研究的meta分析与NMA结合到DTA研究的网络meta分析中。

This project refers to two active research areas of evidence synthesis in medicine, meta-analysis of diagnostic test accuracy studies and network meta-analysis. The final objective is to combine both areas.In standard approaches of meta-analysis of diagnostic accuracy studies, each study is assumed to contribute one pair of sensitivity and specificity. In the first period of the project, we investigated two more general approaches, which account for multiple thresholds of the underlying biomarker, `Modelling biomarker distributions' and `Averaging ROC curves'.The first approach is based on the idea of estimating the distribution functions of an underlying biomarker within the non-diseased and diseased individuals. Based on a distributional assumption, we estimate the distribution parameters in the two groups applying a linear mixed effects model to the appropriately transformed data. The model accounts for both the within-study dependence of sensitivity and specificity and between-study heterogeneity. We obtain a summary receiver operating characteristic (SROC) curve as well as the pooled sensitivity and specificity at every specific threshold. Furthermore, the determination of an optimal threshold across studies is possible through maximization of the Youden index.In the second approach, originally introduced by Martinez-Camblor (2014), the summary ROC curve is determined as a weighted average of the ROC curves of the primary studies by averaging in `vertical' direction (i.e., averaging sensitivities, conditional on specificity). We extended this method by (1) exchanging the roles of sensitivity and specificity, i.e., averaging specificities conditional on sensitivity (horizontal averaging), and (2) averaging the differences of true positive rates and false positive rates (i.e., the Youden indices) conditional on their sum (diagonal averaging).Another area of research was network meta-analysis (NMA). In four publications, we investigated (1) the relation between NMA and electrical network theory, (2) an alternative method of adjusting for multi-arm studies by appropriately inflating the standard errors, (3) methods of automated visualisation of networks, and (4) frequentist treatment ranking, based on a network-meta-analysis.In the second project period we want to refine and extend all these approaches, with a special focus on knowledge translation. There is a more and more increasing spectrum of advanced methods on the one hand and their restricted accessibility for non-statistical users on the other hand. We want to bridge this gap and continue to follow this aim by writing a new R package for implementation of the `Modelling biomarker distributions' and `Averaging ROC curves' approaches and extending our existing R package netmeta for network meta-analysis, for example by meta-regression. The final objective is to combine meta-analysis of DTA studies and NMA to network meta-analysis of DTA studies.