June 7, 2023

Supporting increased transparency in the use of healthcare algorithms: What it means to be "clear"

Department of Learning Health Sciences (DLHS) Associate Professors,  Karandeep Singh, MD, MMSc and Jodyn Platt, MPH, PhD submitted comments to the Office of the National Coordinator for Health Information Technology (ONC) on the April 18 notice of proposed rulemaking (NPRM) titled: Health Data, Technology, and Interoperability: Certification Program Updates, Algorithm Transparency, and Information Sharing (HTI-1).

Drs. Singh and Platt led a multidisciplinary, multi-institutional team to submit support for increasing transparency in the use of healthcare algorithms. The team made four recommendations on what it would mean to be “clear with the public.”

  •  Place the source attributes identified in § 170.315(b)(11)(vi) for predictive DSIs in the public domain. It is not sufficient to have the attributes only available for “user review (https://www.federalregister.gov/d/2023-07229/p-621). The source attributes should be readily accessible by the public and in the public domain. This recommendation is aligned with Executive Order 13960 Promoting the Use of Trustworthy Artificial Intelligence in the Federal Government, which states that “agencies shall be transparent in disclosing relevant information regarding their use of AI to appropriate stakeholders, including the Congress and the public.”14 Having source attributes in the public domain would help to achieve this goal.
  • Add information about which data sources were used for training and evaluation of the predictive DSIs to the set of source attributes and variables identified in § 170.315(b)(11)(vi). Additionally, place this information in the public domain.
  • Add information about all variables (predictors and outcomes) used in predictive DSIs to the set of source attributes and variables identified in § 170.315(b)(11)(vi). Additionally, place this information in the public domain. All variables should be included because demographic and proxy variables may be important for effective evaluation of bias, discrimination, utility, and quality. In cases where listing all variables may not be possible (such as models using x-ray imaging data), a clear statement to that effect should be made.
  • In response to the question as to whether having this information publicly available would improve public confidence in predictive DSIs by enabling research on source attribute information (https://www.federalregister.gov/d/2023-07229/p-709), we believe that making information about models readily available and easy to find is critical to improving public confidence in predictive DSIs and in enabling research on source attribute information. Research should also include studies that inform evidence-based best practices for communications strategies, such as model cards, product labels, and dissemination, and studies that evaluate the impact of governance on transparency and public trust.

HTI-1 comes three years after the 21st Century CURES Act Final Rule, and is meant to build on progress made to support patients and providers across the spectrum of care.

Link to the full letter with recommendations: https://downloads.regulations.gov/HHS-ONC-2023-0007-0024/attachment_1.pdf

Link to the comment: https://www.regulations.gov/comment/HHS-ONC-2023-0007-0024