Federated Learning for Optimal Decision Making in Radiotherapy Using Panomics Analytics

现代放射治疗（RT）的复杂环境包括来自患者的丰富组合的数据， 具体信息包括：人口统计学、高能剂量的物理特征、随后的特征 图像引导（放射组学）和生物标记（基因组学、蛋白质组学等）的重复应用， 在治疗期之前和/或期间产生，所述治疗期可以跨越几天到几周。这些快速增长 可用和未开发的“泛组学”数据，为大数据分析提供了充足的机会， 这在有希望但高风险的RT程序中尤其如此， 作为立体定向体RT（SBRT），由于早期的临床成功， 缩短高剂量治疗的疾病阶段和社会经济效益。这导致了人们的愿望， 将这些治疗方法应用于更晚期的癌症，然而，与之相关的未知风险 增加的毒性阻碍了其潜力。因此，强大的临床决策支持系统（CDSS）能够 探索复杂的泛组学互动景观的目标是利用已知的原则， 在分次RT之前和期间的治疗反应是迫切需要的。的长期目标 该项目旨在克服与预测不确定性和人机交互相关的障碍， 目前限制了为基于实时响应的适应做出个性化临床决策的能力 从现有的数据来看。为了满足这一需求并克服当前的挑战，我们将开发和 定量评估：（1）用于鲁棒预测的基于联邦图的监督机器学习算法 RT之前和期间的结果;（2）联合深度强化学习，以动态优化治疗 适应;和（3）一个以用户为中心的软件原型RT决策支持使用可扩展的XNAT 平台，更广泛的目标是为成果建模和 我们假设使用高级联邦机器学习 技术和以用户为中心的工具将释放潜力， 受主观经验和启发式规则的限制，将其转化为稳健的、患者特异性的、用户友好的CDSS。这 方法及其相应的软件工具将在肺癌和肝癌的两个临床RT站点中进行测试， 以展示其多功能性并突出相关的人机因素和癌症特定问题。 影响声明：患者特定的大数据现在可以在RT课程之前和/或期间使用， 和个性化治疗的未开发机会。这项研究将克服目前的缺点， 通过调查和开发一种基于人口的方法和数据， 联合的以用户为中心的个性化CDSS，需要集中的数据共享，并测试其性能 奖励但高风险的RT场景。该方法也适用于其他现代癌症治疗方案。

The complex environment of modern radiation therapy (RT) comprises data from a rich combination of patient- specific information including: demographics, physical characteristics of high-energy dose, features subsequent to repeated application of image-guidance (radiomics), and biological markers (genomics, proteomics, etc.), generated before and/or over a treatment period that can span few days to several weeks. Rapid growth of these available and untapped “pan-Omics” data, invites ample opportunities for Big data analytics to deliver on the promise of personalized medicine in RT. This is particularly true in promising but high-risk RT procedures such as stereotactic body RT (SBRT), which have witnessed tremendous expansion due to clinical successes in early disease stages and socio-economic benefits of shortened high dose treatments. This has led to the desire to exploit these treatments into more advanced stages of cancer, however, the unknown risks associated with increased toxicities hamper its potential. Therefore, robust clinical decision support systems (CDSSs) capable of exploring the complex pan-Omics interaction landscape with the goal of exploiting known principles of treatment response before and during the course of fractionated RT are urgently needed. The long-term goal of this project is to overcome barriers related to prediction uncertainties and human-computer interactions, which are currently limiting the ability to make personalized clinical decisions for real-time response-based adaptation in radiotherapy from available data. To meet this need and overcome current challenges, we will develop and quantitively evaluate: (1) federated graph-based supervised machine learning algorithms for robust prediction outcomes before and during RT; (2) federated deep reinforcement learning to dynamically optimize treatment adaptation; and (3) a user-centered software prototype for RT decision support using the extendable XNAT platform, with the broader goal of building a comprehensive real-time framework for outcome modeling and response-based adaption in RT. We hypothesize that the use of advanced federated machine learning techniques and user-centered tools will unlock the potentials to move from current population-based approaches limited by subjective experiences and heuristic rules into robust, patient-specific, user-friendly CDSSs. This approach and its corresponding software tools will be tested within two clinical RT sites of lung and liver cancers, to demonstrate its versatility and highlight pertinent human-computer factors and cancer specific issues. Impact statement: Patient-specific big data are now available before and/or during RT courses, offering new and untapped opportunities for personalized treatment. This study will overcome current shortcomings of population-based approaches and data underuse in current RT practice by investigating and developing a federated user-centered, personalized CDSS with the need for centralized data sharing and test its performance in rewarding but high-risk RT scenarios. The approach is also applicable to other modern cancer regimens.