CRITICAL: Collaborative Resource for Intensive care Translational science, Informatics, Comprehensive Analytics, and Learning

点击翻译按钮获取中文摘要

CRITICAL: Collaborative Resource for Intensive care Translational science, Informatics, Comprehensive Analytics, and Learning Translational research in Artificial Intelligence (AI) has been hindered by the lack of shared data resources with sufficient depth, breadth and diversity. There are very limited EHR datasets freely available to the general research community especially the AI research community through credential-based access. MIMIC dataset is from a single institution that has a fixed and limited racial, ethnic and geographic profile. The eICU dataset is limited in data comprehensiveness (e.g., number of kinds of lab tests ~1/5 of MIMIC), data span (1 year, 2014- 2015), and data variety (e.g., no free text clinical notes) etc. Thus MIMIC and eICU respectively have advantages and disadvantages of data depth and data breadth. The vision of this proposal is to leverage multiple CTSAs with diverse racial, ethnic and geographic profiles in order to develop and evaluate a multi-site de-identified ICU dataset, to facilitate accelerate translational research in AI and deep learning approaches to understand, track, and predict the pathophysiological state of patients. In this project, a group of nationwide CTSA sites will work together to build a new, more inclusive, multi-site dataset that is downloadable from NCATS cloud by researchers with credential-based access. This project will combine the respective advantages of MIMIC (data depth) and eICU (data breadth). The created dataset will include more geographic regions, larger quantities of time-series data, including pre-, during- and post- ICU patient information. This will incorporate not only more patient diversity, but also capture regional population differences and practice variations that could have clinical impact. Aim 1 will develop and provide credentialed access to a multi-site dataset consisting of de-identified discrete outpatient, inpatient, and ICU data for critically ill at respective CTSAs. Aim 2 will create federated access dataset from and develop novel federated learning methods on the part of the multi-site ICU data consisting of unstructured clinical notes or structured data for select group of patients at higher risks of re-identification (e.g., rare disease patients). Aim 3 will develop novel memory-network based meta-learning AI algorithms and use the multi-site dataset to answer concrete and long-standing clinical problems in critical care. Aim 4 will innovatively leverage the library network to develop and disseminate open resources for the research community and develop best practice guidelines for other CTSAs to join the effort. In particular, we aim to support and cultivate the growth of next generation medical AI workforce for research and practice. We aim to establish a large cross-CTSA collaborative data sharing for critical care by leveraging the existing CTSA collaborative networks. With the diversified racial, ethnic and geographic profiles from the above CTSAs, we will be able to support fair and generalizable algorithms for advanced patient monitoring and decision support. The proposed project will provide best practice guidance to and set up exemplary examples for nationwide CTSAs. It will also support the cultivation of next generation medical AI researchers.