exponential consumer health data generation have created a data-rich health care ecosystem. However, this growth in health care data is hampered by the lack of efficient mechanisms for integrating and merging these data beyond their current silos. While there are multiple frameworks and standards in place to help aggregate and achieve sufficient data volume for AI use of data at rest (such as mature health care common data models) and data in motion (such as Health Level Seven International Fast Healthcare Interoperability Resources [HL7 FHIR]), they need wider adoption to support AI tool development, deployment, and maintenance. There continue to be issues of interoperability and scale of data transfers due to cultural, social, and regulatory reasons. Solutions will require the engagement of all relevant stakeholders. Thus, the wider health care community should continue to advocate for policy, regulatory, and legislative mechanisms that improve equitable, inclusive data collection and aggregation, and transparency around how patient health data may be best utilized to balance financial incentives and the public good.

ETHICAL HEALTH CARE, EQUITY, AND INCLUSIVITY SHOULD BE PRIORITIZED

Fulfilling this aspiration will require ensuring population-representative datasets and giving particular priority to what might be termed a new Quintuple Aim of Equity and Inclusion for health and health care (see Figure S-1). Else, the scaling

possible with AI might further exacerbate the considerable existing inequities in health outcomes at a monumental scale. A single biased human or organizational impact is far less than that of global or national AI.

Prioritizing equity and inclusion should be a clearly stated goal when developing and deploying AI in health care. There are many high-profile examples of biased AI tools that have damaged the public’s trust in these systems. It is judicious for developers and implementers to evaluate the suitability of the data used to develop AI tools and unpack the underlying biases in the data, to consider how the tool should be deployed, and to question whether various deployment environments could adversely impact equity and inclusivity. There are widely recognized inequities in health outcomes due to the variety of social determinants of health and perverse incentives in the existing health care system. Unfortunately, consumer-facing technologies have often worsened historical inequities in other fields and are at risk of doing so in health care as well.

THE DIALOGUE AROUND TRANSPARENCY AND TRUST SHOULD CHANGE TO BE DOMAIN- AND USE-CASE DIFFERENTIAL

Transparency is key to building this much needed trust among users and stakeholders, but there are distinct domains with differential needs of transparency. There should be full transparency on the composition, semantics, provenance, and quality of data used to develop AI tools. There also needs to be full transparency and adequate assessment of relevant performance components of AI. However, algorithmic transparency may not be required for all cases. AI developers, implementers, users, and regulators should collaboratively define guidelines for clarifying the level of transparency needed across a spectrum. These are key issues for regulatory agencies and clinical users, and requirements for performance are differential based on risk and intended use. Most importantly, we suggest clear separation of data, algorithmic, and performance reporting in AI dialogue, and the development of guidance in each of these spaces.

NEAR-TERM FOCUS SHOULD BE ON AUGMENTED INTELLIGENCE RATHER THAN FULL AUTOMATION

Some of the AI opportunities include supporting clinicians undertaking tasks currently limited to specialists; filtering out normal or low acuity clinical cases so that specialists can work at the top of their licensure; helping humans address inattention, microaggressions, and fatigue; and improving business process

automation. Ensuring human-centered AI tools includes accepting that human override is important for developing user trust because the public has an understandably low tolerance for machine error and that AI tools are being implemented in an environment of inadequate regulation and legislation. The near-term dialogue around AI in health care should focus on promoting, developing, and evaluating tools that support humans rather than attempting to replace them with full automation.

DEVELOP AND DEPLOY APPROPRIATE TRAINING AND EDUCATIONAL PROGRAMS TO SUPPORT HEALTH CARE AI

In order to benefit from, sustain, and nurture AI tools in health care we need a thoughtful, sweeping, and comprehensive expansion of relevant training and educational programs. Given the scale at which health care AI systems could change the medical domain, the educational expansion must be multidisciplinary and engage AI developers, implementers, health care system leadership, frontline clinical teams, ethicists, humanists, and patients and patient caregivers because each brings a core set of much needed requirements and expertise. Health care professional training programs should incorporate core curricula focused on teaching how to appropriately use data science and AI products and services. The needs of practicing health care professionals can be fulfilled via their required continuing education, empowering them to be more informed consumers. Additionally, retraining programs to address a shift in desired skill sets due to increasing levels of AI deployment and the resulting skill and knowledge mismatches will be needed. Last, but not least, consumer health educational programs, at a range of educational levels, to help inform consumers on health care application selection and use are vital.

LEVERAGE EXISTING FRAMEWORKS AND BEST PRACTICES WITHIN THE LEARNING HEALTH CARE SYSTEM, HUMAN FACTORS, AND IMPLEMENTATION SCIENCE

The challenges in operationalizing AI technologies into the health care systems are countless in spite of the fact that this is one of the strongest growth areas in biomedical research and impact. The AI community must develop an integrated best practice framework for implementation and maintenance by incorporating existing best practices of ethical inclusivity, software development, implementation

science, and human–computer interaction. This framework should be developed within the context of the learning health care system and be tied to targets and objectives. The cost and burden of implementing AI tools should be weighed against use case needs. AI tools should be pursued where other low-or no-technology solutions will not do as well. Successful AI implementation will need the committed engagement of health care stakeholders—leaders, AI developers, AI implementers, regulators, humanists, patients, and families. Health delivery systems should have a robust and mature underlying information technology (IT) governance strategy in place prior to them embarking on substantial AI deployment and integration. Lastly, national efforts should be deployed to provide capacity for AI deployment in lower resource environments where IT and informatics capacities are less robust. Linked to the prior considerations, this would help lower the entry barrier for adoption of these technologies and help promote greater health care equity. Health care AI could also go beyond the current limited biology-focused research to address patient and communal needs, expanding to meaningful and usable access of social determinants of health and psychosocial risk factors. AI has the potential (with appropriate consent) to link personal and public data for truly personalized health care.

BALANCING DEGREES OF REGULATION AND LEGISLATION OF AI TO PROMOTE INNOVATION, SAFETY, AND TRUST

AI applications have an enormous ability to improve patient outcomes, but they could also pose significant risks in terms of inappropriate patient risk assessment, diagnostic error, treatment recommendations, privacy breaches, and other harms. Regulators should remain flexible, but the potential for lagging legal responses will remain a challenge for AI developers and deployers. In alignment with recent congressional and U.S. Food and Drug Administration developments and guidance, we suggest a graduated approach to the regulation of AI based on the level of patient risk, the level of AI autonomy, and considerations for how static or dynamic certain AI are likely to be. To the extent that machine learning–based models continuously learn from new data, regulators should adopt postmarket surveillance mechanisms to ensure continuing (and ideally improving) high-quality performance. Liability accrued when deploying AI algorithms will continue to be an emerging area as regulators, courts, and the risk-management industries deliberate. Tackling regulation and liability among AI adopters is vital when evaluating the risks and benefits. Regulators should engage stakeholders and experts to continuously evaluate deployed clinical AI for effectiveness and

safety based on real-world data. Throughout that process, transparency can help deliver better-vetted solutions. To enable both AI development and oversight, government agencies should invest in infrastructure that promotes wider, ethical data collection and access to data resources for building AI solutions within a priority of ethical use and data protection (see Figure S-2).

CONCLUSION

AI is poised to make transformative and disruptive advances in health care. It is prudent to balance the need for thoughtful, inclusive health care AI that plans for and actively manages and reduces potential unintended consequences, while not yielding to marketing hype and profit motives. The wisest guidance for AI is to start with real problems in health care, explore the best solutions by engaging relevant stakeholders, frontline users, and patients and their families—including AI and non-AI options—and implement and scale the ones that meet our Quintuple Aim: better health, improved care experience, clinician well-being, lower cost, and health equity throughout the health care system and all forms of health care delivery.