7
ADVANCING THE AI CODE OF CONDUCT FRAMEWORK: THE “VITAL FEW” PRIORITY ACTIONS

To realize the benefits and avoid the risks associated with health artificial intelligence (AI), it is imperative strategically to prioritize key actions that are most likely to ensure that incentives and supports are intentionally designed and properly executed, and that progress is both effective and responsive to a changing environment. This will require significant effort from and coordination between all stakeholder groups. In addition, beyond the targeted efforts outlined below, much work will be required by multiple parties. The actions listed are not intended to be comprehensive but instead constitute the “vital few” highest priorities to advance the Code Principles and Commitments, building the capability to rapidly respond to new tools, technologies, opportunities, and concerns.

In identifying priority actions for the translation of the Code Commitments to real-world application, information synthesis, gaps, and opportunities were identified by all contributors to this work. Two additional constructs were considered; the first was the identification of priority actions that are foundational to, supportive of, or capable of catalyzing additional needed action, thereby likely to create a cascading effect and potentially speeding the national collective effort to promote safe, effective, and trustworthy health AI. The second was the application, as appropriate, of behavioral economics, which posits a set of strategies including incentives, defaults, and framing to make preferred decisions and actions easier and more rewarding or less costly (Siegel et al., 2021). Below, for each commitment, priority actions are outlined for consideration in the application of the Code Commitments in real-world settings.

COMMITMENT 1: ADVANCE HUMANITY

To ensure that human health, agency, and connection remain the primary focus of health AI, it is essential to identify, transparently characterize, and promote the societal and cultural goals of the recipients of the use of health AI as an accountable mechanism to protect and advance human health and connection.

Development of standards and other governance structures to assess alignment by developers and users of health AI with societal and cultural goals for health AI:

• Requirements to understand end user and recipient needs and preferences.
• Methods at all governance levels to assess the degree to which innovation maintains a foundation of beneficence and the avoidance of harm.
• Patients and patient advocacy groups as central, engaged stakeholders with agency to ensure that this group has the capacity to influence decisions and outcomes in all aspects of health AI.

Incentives and structures for independent evaluation, certification to the AI Code Commitments, and public and transparent reporting on certification status:

• Proactive information disclosure to intended audiences regarding the development and use of AI, actively involving them in the planning process, educating them, and aligning with their needs and preferences.

COMMITMENT 2: ENSURE EQUITY

To ensure equitable distribution of benefits and risks of health AI, it will be critical to place equity and fairness at the center of all health AI development and use and ensure its prioritization throughout the AI lifecycle.

Standardized metrics to assess and report bias in data, AI output, and AI use in the interest of equitable distribution of benefit and risk:

• Diverse, representative data sources for AI training and local implementation.
• Proactive and reactive bias mitigation.
• Metrics assessing fairness and action based on the results.

Incentives and support for low-resourced organizations and communities to ensure equitable access to the benefits of AI:

• Incentives and supports akin to those provided to drive electronic health record adoption.
• Alignment of incentives and payment models to promote equity and democratization of health AI.

COMMITMENT 3: ENGAGE IMPACTED INDIVIDUALS

To ensure that key stakeholders are viewed as partners with agency in every stage of the lifecycle, it is important to identify, engage, and most importantly, integrate all relevant stakeholder input in conceptualization, design, development, implementation, and surveillance throughout the health AI lifecycle.

Participation by all key stakeholders across the health AI lifecycle:

• Engagement with intended users of health AI to ensure preferences and workflow integration where possible.
• Engagement of ethics and equity experts in all AI development.
• Expanded stakeholder inclusion in developing research program funding and goals, which federal agencies and other funding entities could consider.

Local governance bodies that include all stakeholders in the AI lifecycle:

• Local governance frameworks and maturity models.
• Transparent processes for decision making, conflict management, and prioritization to review, understand, and redress stakeholder conflicts, using the Code Principles and Code Commitments as guidance.

Common understanding/education of all impacted parties:

• Patient and end-user education programs designed in collaboration with them, and in a culturally appropriate fashion to support awareness of risks and benefits of health AI and support their personal decision making about its use in their process of care.
• Appropriate communication channels to reach patients to promote a shared understanding of what AI is, what the risks and benefits of AI use are, how it is used for their care, and how they can access available information on AI transparency (akin to ClinicalTrials.gov).

COMMITMENT 4: IMPROVE WORKFORCE WELL-BEING

Consistent with the priorities laid out in the National Academy of Medicine Plan for Health Workforce Well-Being (NAM, 2022c), it is imperative to create a shared sense of purpose and potential for the health care workforce. Top priorities include workforce education and investment in research.

Positive work and learning environments and culture (NAM, 2022c):

• Curriculum standards for trustworthy AI competencies across the spectrum of stakeholders including developers, researchers, and the health care workforce.
• Well-defined communication and educational programs to support health care workforce AI competence and comfort in use.
• Programs and evaluation mechanisms for workforce reskilling to facilitate retention, adaptation of roles, and implementation of change management processes to ease the introduction of disruptive technologies.
• Training and expansion of the health care workforce that is AI aware and AI competent.

Measurement, assessment, strategies, and research (NAM, 2022c):

• Advancement of science and understanding of the interaction of AI technologies and health care delivery workflow.
• Change management tools and techniques to develop capacity for workflow redesign.

COMMITMENT 5: MONITOR PERFORMANCE

Effective monitoring and sharing of AI’s performance and impact on health and safety will require stakeholders to integrate and align risk management and quality measurement and assurance frameworks for the health AI lifecycle. Careful consideration is needed to assess technical rigor, use case utility, and trustworthiness (equity, fairness) in the conduct of performance monitoring.

Standardized quality and safety metrics to assess the impact of the use of health AI on health outcomes:

• Processes that can document and report on technical and clinically meaningful performance metrics that are understandable to the public, patients, and their caregivers.

• Requirements and frameworks regarding health AI performance monitoring that balance safety, health utility, and flexibility.

COMMITMENT 6: INNOVATE AND LEARN

Innovation and discovery are needed to drive continuous improvements to health, and shared learning and ongoing systems feedback are the foundation of the Learning Health System.

A well-supported national health AI research agenda:

• Federal research, development, and implementation efforts to support an ecosystem of safe and effective health AI.
• Collaborative research projects involving health care providers, patients, AI developers, and academic institutions that explore AI applications and their implications in clinical settings.
• Efforts to assess not only development and implementation but also the assessment of decommissioning activities.

Participation in shared learning across all stakeholders:

• Meetings, workshops, or conferences to exchange insights, recognize shared obstacles, and create joint solutions.

Innovation as a core investment:

• Entrepreneurs, start-ups, and small businesses that deliver novel solutions to problems in health, health care, and biomedical science.
• Investment in research and advancement in science among federal agencies for critical needs in health AI where commercial stakeholders may have less focus.
• Ongoing support for public–private partnerships as a means of promoting innovation in health AI.

While there is clearly much work to be done across stakeholders to advance responsible health AI, prioritizing actions that impact the highest points of leverage and that are in some cases already in motion will allow us to reap the benefits and avoid the pitfalls of health AI most expeditiously, safely, and effectively. Applying the concepts of behavioral economics, it will be important to make it easy and rewarding to abide by the shared vision, values, goals, and expectations described

in the nationally aligned AI Code Principles and Code Commitments. Established standards, incentives, and transparent performance metrics will be key. Table 7-1 summarizes priority actions to operationalize the AICC Code Commitments.

TABLE 7-1 | Summary of Priority Actions to Operationalize the AICC Code Commitments

NOTE: ASTP ONC = Assistant Secretary for Technology Policy, Office of the National Coordinator for Health Information Technology; FDA = U.S. Food and Drug Administration; HRSA = Health Resources and Services Administration; NIH = National Institutes of Health.

This page intentionally left blank.