A System for Xerostomia Risk Classification after Head & Neck Cancer Radiotherapy

摘要 放射治疗(RT)是大多数头颈癌(HNC)治疗的主要组成部分。在.期间 辐射后，唾液腺等敏感部位会受到伤害，导致口干症(口干)。 这种副作用很常见，会显著降低治疗过程中和治疗后的生活质量。的关注点 这一应用是在治疗计划中预测患者是否会遭受严重口干症。 (NCI CTCAE 2-3级)在第一次治疗后随访时，通常在RT后3-6个月 (患病率约为40%)。预测将使临床医生能够进行治疗计划 提高对高度口干症发生的可能性的认识，并允许更多的信息和更多的信息 及时预见进食困难等后果。 在这个第一阶段的项目中，OncSpace Inc.将开发一个分类和回归树(CART)预测 模型使用了1200多个完整的HNC患者记录。重度口干症与广泛性口炎之间的关系 将自动发现一系列剂量学、临床和人口学特征，并使用 最强关联将填充决策树的节点。每个末端叶节点都将包含 该结节中的患者亚组发生高度口干症的概率。此外，还将分配叶节点 标明高风险或低风险的高度口干症的二元分类标签。这种类型的模型提供了 透明度和可解释性，这有利于临床接受和证明 监管机构。该软件将使用Microsoft Azure云架构构建，并通过 软件即服务(SaaS)模型。 这个项目有三个不同的目标： 1.用约翰·霍普金斯大学授权的数据填充OncSpace Inc.的S Microsoft Azure CosmosDB数据库 大学，包括取消患者身份识别、数据管理和其他数据集功能等步骤 工程学 2.使用单独的训练和测试数据集以及各种 包括敏感度、特异度、AUC和F1得分。 3.设计了临床可接受的风险分类策略和用户界面(UI)来交流模型 结果。来自UI顾问团队和三名放射肿瘤学家的专家意见将是不可或缺的部分 开发、测试和评估过程。 这些目标的成功实现将证明口干症的临床和商业可行性。 HNC的预测模型。第二阶段的进一步开发将包括通过以下方式进行更深入的模型个性化 纳入高级图像特征(放射组学)，以及验证模型的泛化能力和 通过对来自其他机构的数据进行管理和用于建立模型来实现商业可行性。 成立于2018年的OncSpace处于开展这项工作的独特地位，因为该团队包括 顶峰放射治疗计划系统、放射治疗放射治疗传送系统和HealthMyne 定量成像决策支持平台。OncSpace与约翰·霍普金斯大学有着密切的临床合作 用于临床反馈、验证和初步部署的JHU。OncSpace已经授权了三项专利 和订阅JHU超过6,000名放射肿瘤患者的完整患者治疗记录。这个 该公司凭借其创新平台获得了微软创新加速奖，以提供支持AI的 为放射肿瘤学社区提供医疗保健解决方案。 第1页，共1页

Summary Radiotherapy (RT) is a major component in the treatment of most head and neck cancer (HNC) cases. During irradiation, sensitive regions such as the salivary glands can sustain injury, resulting in xerostomia (dry mouth). This side effect is common and can significantly reduce quality of life during and post-treatment. The focus of this application is prediction during treatment planning of whether patients will suffer high-grade xerostomia (NCI CTCAE Grade 2-3) at the time of their first post-treatment follow-up visit, typically 3-6 months after RT (prevalence is approximately 40%). Predictions will enable clinicians to carry out treatment planning with improved knowledge of the likelihood of high-grade xerostomia development and allow better-informed and more timely anticipation of consequences such as eating difficulty. In this Phase 1 project, Oncospace Inc. will develop a Classification and Regression Tree (CART) prediction model using over 1200 complete HNC patient records. Associations between high-grade xerostomia and a wide range of dosimetric, clinical and demographic features will be automatically discovered and the features with the strongest associations will populate the nodes of a decision tree. The terminal leaf nodes will each contain the probability of high-grade xerostomia for the subset of patients in that node. In addition, leaf nodes will be assigned binary class labels designating a high- or low risk of high-grade xerostomia. This type of model provides transparency and interpretability, which are beneficial for clinical acceptance and for demonstration of safety to regulatory agencies. The software will be built using the Microsoft Azure cloud architecture and be deployed via a Software as a Service (SaaS) model. There are three distinct aims of this project: 1. Populate Oncospace Inc.’s Microsoft Azure CosmosDB database with data licensed from Johns Hopkins University, including steps such as patient de-identification, data curation, and additional dataset feature engineering 2. Perform CART modeling and test model accuracy, using separate training and test datasets and a variety of performance metrics, including sensitivity, specificity, AUC, and F1-score. 3. Design a clinically acceptable risk classification strategy and a user interface (UI) to communicate model results. Expert input from a team of UI consultants and three radiation oncologists will be an integral part of the development, testing, and evaluation processes. The successful completion of these aims will demonstrate the clinical and commercial feasibility of a xerostomia prediction model for HNC. Further development in Phase 2 will include deeper model personalization via incorporation of advanced image features (radiomics), as well as validation of model generalizability and commercial viability via the curation and use in model building of data from other institutions. Oncospace, formed in 2018, is uniquely positioned to carry out this work as the team includes the creators of the Pinnacle radiation therapy planning system, Tomotherapy radiation treatment delivery system, and HealthMyne Quantitative Imaging Decision Support platform. Oncospace has close clinical collaboration with Johns Hopkins University (JHU) for clinical feedback, validation and initial deployment. Oncospace has licensed three patents and subscription to complete patient treatment records for over 6,000 radiation oncology patients from JHU. The company has won the Microsoft Innovation Acceleration Award for its innovative platform to deliver AI-enabled healthcare solutions to the radiation oncology community. Page 1 of 1