Bridge2AI: Patient-Focused Collaborative Hospital Repository Uniting Standards (CHoRUS) for Equitable AI

迫切需要基础设施来支持重症监护中的人工智能和机器学习(AI/ML)。开发高分辨率多中心数据集是迈向可操作和值得信赖的人工智能的关键第一步。作为NIH共同基金Bridge2AI计划的一部分，针对公平AI数据生成的以患者为中心的协作医院存储库统一标准(CHOUS)项目将满足为ML/AI应用程序生成数据的需求，该应用程序旨在表征急性和危重疾病、预测并发症和衡量急危重症患者的治疗反应。通过6个模块，以患者为中心的公平人工智能数据生成项目将解决从100,000多名危重患者中获取人工智能就绪数据集的多个相关挑战：1)团队科学，2)道德和值得信赖的人工智能，3)标准，4)工具开发和优化，5)数据获取，以及6)技能和劳动力发展。该项目的总体目标是开发一个公开可用的、为人工智能做好准备的前所未有的多样性重症监护数据集，同时确保这些方法促进隐私、问责、临床利益和公平，同时促进新一代人工智能临床医生和科学家。该数据集还将包括一个坚持测试集，可供模型外部验证访问，以帮助市场采用人工智能开发的模型，以便在急性和危重护理中实施。 该项目从团队科学、法律、伦理学、卫生服务、生物医学、工程学和科学期刊出版物等不同学科中汲取专业知识，将A)建立一个大规模收集数据的法律框架，抽样以确保多样性和最大限度地减少偏见；B)开展面向社区的伦理焦点小组，以确定哪些数据适合于公共共享；C)确保数据元素包括适当的健康社会决定因素，以研究和了解保健提供方面的潜在偏见；D)开发跨多中心的能力，以获取、标准化、标记化、存储、可视化和标记数据，包括结构化电子健康记录数据、标记化非结构化电子健康记录数据、遥测和脑电波形、成像和健康的社会决定因素；e)获取数据，将数据标准化到OMOP通用数据模型，使用限制重新识别的不同隐私方法转换数据，并为诊断和临床恶化事件标记数据；以及f)培养外行和科学界的专业知识，通过多模式教育方法提高人工智能素养和利用率。为了实现这一目标，该项目将涉及中心之间以及通过NIH Bridge2AI计划、NIH Bridge2AI Bridge Center、外部生物医学和临床组织、行业和监管机构的广泛合作。

There is an urgent need for infrastructure to support artificial intelligence and machine learning (AI/ML) in critical care. Developing high-resolution multi-center data sets is a critical first step towards actionable and trustworthy AI. As part of the NIH Common Fund’s Bridge2AI program, the Patient-Focused Collaborative Hospital Repository Uniting Standards (CHoRUS) for Equitable AI data generation project will meet the need of generating data for ML/AI applications aimed at characterizing acute and critical care illness, predicting complications, and measuring treatment response among patients with acute or critical illness. Through 6 modules, the Patient-Focused CHoRUS for Equitable AI data generation project will addresses multiple challenges relevant for acquiring an AI-ready data set from more than 100,000 critically ill patients: 1) Team Science, 2) Ethical and Trustworthy AI, 3) Standards, 4) Tool Development and Optimization, 5) Data Acquisition, and 6) Skill and Workforce Development. The project’s overarching goal is to develop a publicly available, AI-ready critical care dataset of unprecedented diversity, while ensuring the methods promote privacy, accountability, clinical benefit, and equity, while promoting a new generation of AI clinicians and scientists. The dataset will also include a holdout test set, accessible for model external validation to aid marketplace adoption of AI-developed models for implementation in acute and critical care. Drawing expertise from a diverse range of disciplines including team science, law, ethics, health services, biomedical science, engineering, and scientific journal publications, this project will A) establish a legal framework for collecting data at scale, sampling to ensure diversity and minimize bias; B) perform community-facing ethics focus groups to determine what data is appropriate for public sharing; C) ensure that data elements include appropriate social determinants of health to study and understand potential bias in care delivery; D) develop capabilities across a multi-center to acquire, standardize, tokenize, store, visualize, and label data including structured electronic health record data, tokenized unstructured electronic health record data, telemetry and EEG waveforms, imaging, and social determinants of health; E) acquire data, standardize data to the OMOP Common Data Model, transform data using differential privacy approaches that limit re-identification, and label data for diagnoses and events of clinical deterioration; and F) cultivate expertise in the lay and scientific community to improve AI literacy and utilization through multimodal educational approaches. To accomplish this, the project will involve extensive collaboration between centers as well as through the NIH Bridge2AI program, the NIH Bridge2AI Bridge Center, external biomedical and clinical organizations, industry, and regulatory agencies.