Using Synthetic Controls to Improve Randomised Control Trials for Rare

由于缺乏参与者、伦理和成本，执行罕见疾病治疗的随机对照试验通常很困难（Thorlund et al., 2020）。在这些随机对照试验中，合成对照组有可能取代安慰剂或对照组。本研究有两个目的：首先，确定哪种类型的历史数据在生成合成控制时是可以接受的，以及不同方法之间的准确性差异是什么；其次，创建一个机器学习模型，可以预测给定历史实验臂的控制臂。对于第一个目标，将使用IPD（个体参与者数据）的历史RCT试验。将根据三种不同类型的历史数据（既往rct、观察性研究数据和外部数据）生成单独的合成对照组，并与真实的历史对照组进行比较。对于第二个目标，基于公开可用的摘要级RCT数据的控制臂和实验臂的模拟数据集将用于训练机器学习模型。历史随机对照试验也将从所有可用的来源中招募，用于模型的检验。在第一个目标中，现有的软件包将用于从先前的随机对照试验、观察性研究和外部数据中生成合成对照。使用Pearson相关系数将合成对照与历史对照和实验对照进行比较。在第二个目标中，将使用生成对抗网络（GAN）深度学习模型从输入数据集生成控制数据。生成的数据将使用Pearson相关系数与测试数据进行比较。

It is often difficult to execute a RCT for a rare disease treatment due to lack of participants, ethics, and cost (Thorlund et al., 2020). Synthetic control arms could potentially be used in place of placebo or control arms in these RCTs. There are two aims of this study: first, to determine which types of historical data are acceptable in generating synthetic controls and what the difference in accuracy is between the methods, and second, to create a machine learning model that can predict control arms given historical experimental arms.For the first aim, historical RCT trials with IPD (individual participant data) will be used. Separate synthetic control arms will be generated based on three different types of historical data (previous RCTs, observational study data, and external data) and compared to the real historical control arm. For the second aim, simulated datasets of control arms and experimental arms based on publicly available summary level RCT data will be used to train the machine learning model. Historical RCTs will also be recruited from all available sources for the testing of the model.In the first aim, existing packages will be used to generate synthetic controls from previous RCTs, observational studies, and external data. Synthetic controls will be compared to historical control arms and experimental arms using Pearson's correlation coefficient. In the second aim, a generative adversarial network (GAN) deep learning model will be used to generate control data from input dataset. The generated data will be compared to the test data using Pearson's correlation coefficient.