Deep Learning Approaches to Risk Stratification in Acute Gastrointestinal Bleeding

项目总结 急性胃肠道出血造成超过220万天的住院日和192亿美元的 每年的医疗费用。52%至55%的急性胃肠道出血患者是不必要的 住院，导致资源浪费。尽管出现胃肠道症状的患者的风险分层 建议出血，风险评估评分系统在实践中并不常用，有分 最佳性能，可能应用不正确，并且不容易更新。 大多数当前的风险评分都是根据出血源的位置设计的：上部或 下胃肠道。然而，出血源的位置在出现时并不总是很清楚。一个 根据出现的症状(如呕血、黑便、便血)进行初步评估的风险分数为 在临床实践中更具相关性和实用性。电子健康记录可以用来识别患有 急性胃肠道出血症状并提取临床数据以自动计算风险分数 向供应商提供。机器学习(使计算机能够在没有学习的情况下进行学习的研究领域 被显式编程)，特别是使用神经网络的深度学习(处理 和传输信息)，可以创建比临床风险更好的基于电子健康记录的模型 胃肠道出血评分，非常适合从新数据中学习。 这项提议将使用电子健康记录数据的深度学习工具来减少不必要的 急性胃肠道出血患者的住院治疗通过识别低危患者。目标是1) 开发并验证用于初始风险分层的准确且具有临床实用价值的深度学习算法 现有的临床风险评分2)开发和验证动态深度学习工具以模拟随时间推移的风险，以及 3)在电子健康记录中试行性能最好的算法。将开发深度学习算法 使用来自两个国家的7,000名急性消化道出血患者的电子健康记录数据集 耶鲁-纽黑文卫生系统中的学术医院。将对单独的数据集执行验证 位于马萨诸塞州波士顿的Partners Healthcare的患者。神经网络方法将应用于 患者的数据随着时间的推移而更新，以评估需要输注红细胞的轨迹。最后， 一项试验性研究将在电子健康记录中实施性能最好的算法，为期3个月 以测试部署的可行性和提供者和患者的接受性。计划中的课程包括深度 利用生物医学数据学习、风险评估和纵向分析。 这项工作有可能通过更好地对患者进行综合风险分层来节省成本 出现明显的胃肠道出血。为了达到这项建议的研究和教育目标， 导师团队包括一名胃肠道出血的主要导师和深度学习的联合导师， 基于电子健康记录的临床试验预后算法设计，并实现科学。

PROJECT SUMMARY Acute gastrointestinal bleeding accounts for over 2.2 million hospital days and 19.2 billion dollars of medical charges annually. 52% to 55% of patients with acute gastrointestinal bleeding are unnecessarily hospitalized, leading to wasted resources. Although risk stratification of patients presenting with gastrointestinal bleeding is recommended, risk-assessment scoring systems are not commonly used in practice, have sub- optimal performance, may be applied incorrectly, and are not easily updated. Most current risk scores were designed for use based on the location of the bleeding source: upper or lower gastrointestinal tract. However, the location of the bleeding source is not always clear at presentation. A risk score that bases initial assessment on presenting symptoms (e.g., hematemesis, melena, hematochezia) is more relevant and useful in clinical practice. The electronic health record can be used to identify patients with acute gastrointestinal bleeding symptoms and extract clinical data to automatically calculate risk scores that are made available to providers. Machine learning (field of study that gives computers the ability to learn without being explicitly programmed), particularly deep learning using neural networks (collection of nodes that process and transmit information), can create electronic health record-based models that perform better than clinical risk scores for gastrointestinal bleeding and are well-suited for learning from new data. This proposal will use deep learning tools on electronic health record data to decrease unnecessary hospitalization in patients with acute gastrointestinal bleeding by identifying low risk patients. The goals are to 1) Develop and validate an accurate and clinically useful deep learning algorithm for initial risk stratification superior to existing clinical risk scores 2) Develop and validate a dynamic deep learning tool to model risk over time, and 3) Pilot the best performing algorithms in the electronic health record. Deep learning algorithms will be developed using a dataset of electronic health record data of 7,000 patients with acute gastrointestinal bleeding from two academic hospitals in the Yale-New Haven Health System. Validation will be performed on a separate dataset of patients at Partners Healthcare in Boston, Massachusetts. Neural network approaches will be applied to patients’ data updated over time to evaluate the trajectory towards requiring transfusion of red blood cells. Finally, a pilot study will implement the best-performing algorithms in the electronic health record for a 3-month period to test feasibility of deployment and acceptability to providers and patients. Planned coursework includes deep learning with biomedical data, risk assessment and longitudinal analysis. This work has potential to generate cost savings through better integrated risk stratification of patients presenting with overt gastrointestinal bleeding. To meet the research and educational goals of this proposal, the mentorship team includes a primary mentor in gastrointestinal bleeding and co-mentors in deep learning, electronic health record-based clinical trial design of prognostic algorithms, and implementation science.