Harmonizing multiple clinical trials for Alzheimer's disease to investigate differential responses to treatment via federated counterfactual learning

治疗阿尔茨海默病(AD)的药物开发一直具有挑战性和昂贵。 药物失败很可能在很大程度上是由于患者对 不同的治疗方法。一些亚组患者使用了治疗调节剂并有反应 不同的。由于样本量有限，识别这样的响应子集一直是一项挑战 在一项临床试验中或可能超出个别临床特别分析的范围 试验，考虑到AD的复杂性。另一个重要的患者亚群是快速 进步者，他们在规定的时间内认知下降的速度更快，并可能做出反应 与其他AD患者的治疗不同。预测快速进步者及其 不同的反应非常具有挑战性。机器学习的预测并不比 由于认知分数的波动和综合测试的不足而造成的随机猜测 细粒度的纵向临床数据。汇集来自多个临床试验数据的患者级别数据 可以通过增加样本量和获得更好的 患者群体的覆盖面/代表性。然而，许多临床试验数据被存储 在分布式数据访问服务器中，数据使用协议通常禁止导出患者- 将数据从本地服务器升级。我们的目标是通过先进的信息学来应对这些挑战 使用AI/ML模型的工具。我们将开发保护隐私的联邦模型来协调 将局部反事实效应估计模型转换为全局模型，而无需交换患者- 标高数据。目标1致力于开发一种基于差分的联邦子分组模型 回应。目标2专注于使用以下工具开发联邦反事实回归模型 深度学习预测快速进步者及其差异反应。目标3侧重于 在全国范围内利用真实世界观测验证和提炼亚群预测 财团数据。如果成功，该项目将有助于确定患者亚组 应对方式不同，这将使AD临床规模更小、成本更低、针对性更强 让更少的患者接触到他们不太可能接受的实验性药物的试验 请回答。

Drug development for treating Alzheimer's disease (AD) has been challenging and expensive. Drug failures are very likely due, in large part, to the differential responses of patients to different treatments. Some subsets of patients have treatment moderators and respond differently. Identifying such responsive subsets has been challenging due to limited sample size in one clinical trial or may be beyond the scope of the ad-hoc analyses in individual clinical trials, considering the complexity of AD. Another important subset of patients are rapid progressors, who have faster rates of cognitive decline in a defined period and may respond differently to treatments than other AD patients. Predicting the rapid progressors and their differential responses is very challenging. Machine learning prediction has been no better than random guesses due to volatility of cognitive scores and insufficiency of comprehensive and fine-grained longitudinal clinical data. Pooling patient-level data from multiple clinical trials data may address the above challenges by increasing sample size and obtaining a better coverage/representation of the patient population. However, many clinical trials data are stored in distributed data access servers, and data use agreements often prohibit exporting the patient- level data out of the local servers. We aim to address the challenges via advanced informatics tools using AI/ML models. We will develop privacy-preserving federated models to harmonize local counterfactual effect estimation models into a global model without exchanging patient- level data. Aim 1 focuses on developing a federated subgrouping model based on differential responses. Aim 2 focuses on developing a federated counterfactual regression model using deep learning to predict rapid progressors and their differential responses. Aim 3 focuses on verifying and refining the subgroups prediction using real-world observation in nation-wide consortium data. If successful, this project will contribute to identifying patient subgroups that respond differently, which will result in smaller, less expensive, and more targeted AD clinical trials that expose fewer patients to experimental medications to which they are unlikely to respond.