CHoRUS Executive Summary

Researchers have made great strides in improving patient care through the use of artificial intelligence (AI) in critical care environments such as intensive care units (ICUs). Studies have yielded many positive results from using AI in critical care (AICC)—for example, early detection of organ failure and arrhythmia as well as using AI to assess vital signs and predict potential health risks for admitted hospital patients.

Despite these recent advances, three problems must be solved for AICC research to continue advancing in the pursuit of improving patient care and saving lives:

  1. Lack of a large, diverse data set that can be accessed for research and treatment of patients from various geographical, socioeconomic, and cultural backgrounds.
    Existing large, high-resolution data sets are available only at single centers and are specific to certain demographics or regions, so those data sets are not available or practical for general, widespread use.
  2. Lack of standards that support interoperability (i.e., exchange and use of health information) between different sites and researchers.
    Waveform data—such as ECG, EEG, and blood pressure data—are rarely recorded in sufficiently detailed samples. Privacy protection standards vary between sites nationally, and demographic information may be lost due to those varying privacy standards. Without unified standards to ensure both data integrity and patient privacy, current decisions about data formats are being made locally, which prevents interoperability.
  3. Lack of a trained AI workforce and a public familiar with medical AI.
    To progress in creating an AICC ecosystem that can effectively engage in AI research and improve AI algorithms, the absence of a trained medical AI workforce must be addressed. To make patients comfortable with the use of AI diagnostics and monitoring, the general public needs familiarity with medical AI and accurate information about its potential benefits.

Unlocking the potential of medical AI

CHoRUS—a consortium of 18 top medical research institutions*—received a National Institutes of Health (NIH) Bridge2AI grant to solve those three problems by creating the following:

  1. A 100,000-patient data set with health data from a geographically, socioeconomically, and culturally diverse sample population.
    Researchers and clinicians will finally have a comprehensive data set with the geographic and demographic diversity needed for model generalization. The CHoRUS data set will enable researchers to account for sociodemographic variables and social determinants of health, including a composite index of 17 census measures. The CHoRUS team of AI ethics experts will examine the data set and help to resolve any inequities.
  2. Standards and AICC tools to facilitate interoperability between different ICUs and different researchers.
    A unified set of standards will establish best practices for how to record AI data, what file format to use, and how to structure electronic health record files to incorporate AI data while protecting patient privacy. These tools and standards will allow interoperability and encourage AI research.
  3. Training to expand the AICC research workforce, and public engagement to familiarize general audiences with medical AI.
    To help build a trained, widespread medical AI workforce, the CHoRUS project will conduct workshops for early- and mid-career researchers that include those from disadvantaged backgrounds. Public engagement efforts and AI education modules will provide general audiences with clear and accurate information on AICC research, which could help to resolve personal concerns over the implementation of AI in healthcare.

Beyond the four-year span of the CHoRUS project (2022–2026), these accomplishments will continue to yield benefits for researchers and, ultimately, patients. The 100,000-patient data set, the establishment of a more robust AICC ecosystem, and the trained medical AI workforce will serve the public by providing better critical care to patients in all regions and of all backgrounds.

Expanded AI capabilities and access for patients

High-resolution data in context is key to unlocking early-detection and diagnostic tools for managing sepsis, seizures, cardiac arrest, heart failure, lung injury, and intracranial pressure. The CHoRUS project’s standardization of this high-resolution data is essential for the expansion and use of AICC.

Furthermore, the development of user-friendly software tools will make the data set easier for researchers and clinicians to access and interpret. The CHoRUS team will create an Integrated Viewer application to provide access to all the available data in an integrated fashion (Figure 1).

Example of a multimodal dashboard.
Figure 1. Example of a multimodal dashboard.

The accessibility of the data set will be complemented by the CHoRUS project’s skills and workforce development efforts, which are aimed at expanding the AI workforce in healthcare and establishing a broader range of people in that workforce. These workforce development efforts will include the creation of the undergraduate course “Medical AI, Culture, and Society” as well as a series of short tutorials about medical AI that will be publicly available on the CHoRUS website.

Equitable and ethical AI

Current standards do not always result in both (a) the privacy of patient information and (b) the demographic detail and diversity required for an effective data set. To ensure the privacy of patient data, a team of legal, ethical, and communications scholars will establish a set of recommended guidelines above and beyond the current legal requirements for privacy protection for medical AI research. To ensure the demographic diversity that is required for a data set to be useful for a large and diverse geographical area, the CHoRUS project will use the following avenues:

  • Public engagement with the Citizen Scientist community group and other community groups to learn how to make a broader range of people comfortable with medical AI in their treatment.
  • Workforce training events for beginning and experienced medical AI staff, including those from disadvantaged backgrounds.
  • An Ethical and Trustworthy AI (ETAI) team, to confirm that the CHoRUS data set will benefit patients and researchers in many different geographic areas.

Long-term benefits

The CHoRUS project’s 100,000-patient data set will provide a general model that can be used for AICC research in a larger geographical area and in treatments of people from all backgrounds. The workforce development will encourage AICC research by expanding the medical community’s familiarity with AI tools. The standardization of high-resolution AI data will enable interoperability among AICC researchers and will ensure that patient privacy is preserved. With an expanded AICC ecosystem and greater confidence in its use, AICC research can become more ubiquitous in use and research that will improve recovery from acute illness.

*CHoRUS Institutions

University of Florida, Massachusetts General Hospital at Harvard, Beth Israel Deaconess Hospital at Harvard (through Massachusetts Institute of Technology), Columbia University, University of Pittsburgh Medical Center, Duke University, University of Virginia, UCLA, Emory University, Nationwide Children’s Hospital, Mayo Clinic, Seattle Children’s Hospital, University of New Mexico, UC San Francisco, Johns Hopkins Medical Center, Tufts University Medical Center, University of Oregon, and University of Texas Health Science Center, Houston.

CHoRUS team