AI Governance in Healthcare

Introduction to AI Governance

Building robust governance architectures that create trust in AI/ML technologies is essential for advancing health information governance in healthcare, focusing on transparency, accountability, fairness, and patient safety to improve clinical outcomes and reduce health disparities, as recommended by the Circulation journal 1
Key goals of AI governance in healthcare include empowering patients, ensuring affordable digital health, protecting digital rights, and regulating the digital-health ecosystem, with a focus on underrepresented groups 1
Governance models must adapt to different societal contexts while accounting for implications on individual health and well-being, with a strength of evidence supported by 1

Data used for AI development must be adequate, representative, well-characterized, and reusable, with formal assessments of bias and societal impact conducted and reported, including equality impact assessments, algorithmic impact assessments, and medical algorithmic audits 1, 2
Transparent documentation of datasets is critical for mitigating algorithmic bias and promoting health equity, with key data quality dimensions including accuracy, consistency, security, timeliness, completeness, reliability, accessibility, objectivity, relevancy, and understandability, as identified by the Journal of Medical Internet Research 3, 2
Data users must identify uncertainties or variable performance in groups and clearly state clinical implications as risks, with a focus on ensuring high-quality patient care 2

Strategies to monitor, manage, and reduce risks should be documented as part of AI implementation, with post-market surveillance and clinical follow-up essential, especially when risk of harm differs between groups 2
Decision support systems need regular updates to mitigate effects of changing data quality, population characteristics, and clinical practices, with a strength of evidence supported by 1
Implementation of technical safeguards, including firewalls, secure transmission modes, and encryption, protects electronic protected health information (ePHI) and ensures data quality and privacy protection, as recommended by the American Medical Informatics Association 4, 5

Open-source software improves transparency and participation in AI technology design, with citizen science involving non-professional scientists in research to broaden perspectives 1
Increase diversity of data by promoting involvement of people familiar with potential bias, context, and regulations throughout algorithm development, with a focus on ensuring equitable access to healthcare and improving health outcomes 1
Patients should "own" their data and be able to grant care team members access to it, to ensure patient-centered care and improve outcomes, as recommended by the Journal of the American Medical Informatics Association 5

When necessary, implement debiasing techniques to decrease variation in performance across subgroups, with a focus on reevaluating race correction practices that may exacerbate inequities in disease outcomes and treatments 1
AI governance architectures should engage all stakeholders (developers, clinicians, researchers) to continuously evaluate safety and effectiveness, with companies filing applications with regulatory bodies like FDA to allow marketing of algorithms 1
Post-market safety monitoring similar to phase IV drug trials should be implemented, with a strength of evidence supported by 1

Proactive regulatory approaches are needed to mitigate AI harms before they happen, with a coordinated national approach to protect data more effective than relying solely on health systems and vendors, as recommended by the Journal of the American Medical Informatics Association 5
Standardized methodologies for data quality improvement, such as plan-do-study-act (PDSA) or define-measure-analyze-improve-control (DMAIC) cycles, should be implemented to ensure high-quality patient care, with a focus on addressing coding and semantic variations to ensure data quality across systems 3
Research should be supported to assess alternative approaches to clinical registry linkage and identify the most accurate and efficient methods, as recommended by the American College of Cardiology 6

A multidisciplinary approach involving clinical, technical, and administrative stakeholders is recommended to implement information governance, with regularly reviewing compliance and effectiveness of governance measures, and updating the governance framework based on emerging threats, technologies, and regulations, as suggested by the American Medical Informatics Association 4
Establishment of data sharing review committees with appropriate stakeholder representation, similar to Institutional Review Boards, is recommended by the American Medical Informatics Association 4
Clear definition of roles for data stewards, managers, and users within the organization is suggested by the Clinical Microbiology and Infection society 7
Making the purpose of data sharing activities transparent to all stakeholders, including patients, is proposed by the American Medical Informatics Association 4

Most effective data quality improvement initiatives include multiple interventions, such as DQ reporting and personalized feedback (61% of successful interventions), IT-related solutions (54% of successful interventions), training (44% of successful interventions), improvements in workflows (13% of successful interventions), and data cleaning (8% of successful interventions), as identified by the Journal of Medical Internet Research 3
Considerable heterogeneity exists in settings and approaches to data quality assessment and improvement, highlighting the need for standardized frameworks and approaches, with a strength of evidence supported by 3
Use of standardized data quality frameworks is limited, with most assessments being largely manual, which can negatively impact data quality and patient care, as recommended by the Journal of Medical Internet Research 3