Table 2

Expectations and dependencies

Ensuring accuracy, freedom from bias, trustworthiness.3–5 8 19 20 23 24 29AI applications should be based on models that, in their development, have involved domain experts and have minimised bias related to under-representation of patient groups or contextually inappropriate outcome measures, and have been shown to produce accurate results in the populations for which they are to be used.
Improving efficiency and reduced administrative burden.3–5 10 13–15 17 18,AI applications must be fitted to, and complement, routine clinical workflows and, where possible, self-populate the required data with minimal clinician input.
Improving clinical decision-making and outcomes.3 11 18 21 22 25 27 29AI applications must be shown to be as or more effective in improving clinical decision-making and patient experiences and outcomes than current care, not just efficacious in controlled research settings, and be accompanied with clinician oversight.
Maintaining the integrity of clinician-patient relationships.5 13 18 19 24AI applications should not distract from, or degrade, human to human interaction and shared decision-making.
Ensuring explainability and transparency.16 19 20 23AI applications must be developed and assessed with an eye to maximising explainability and transparency in regards to their inner workings, while acknowledging limits to the extent this can be achieved. As much as possible, important features underpinning AI predictions should be identified, and outputs should be presented in ways easily interpretable to clinicians and patients.
Preserving professional status.3–9 11 12 18AI applications must be implemented with care regarding potential loss of jobs or professional reputation, highlighting the potential of AI to remove the tedious aspects of work, improve job satisfaction and provide new skills. This must be coupled with careful attention to clinicians’ training needs and career development.
Obtaining regulatory approval.3 19 21 27 29AI applications should be subject to regulatory standards that are robust, transparent and responsive to updates of existing applications.
Determining liability for error.3 10 19 21 29AI applications should be associated with clear lines of responsibility regarding liability for error, including no-fault provisions when, despite good evidence of efficacy and safety, errors occur as a result of technical failures involving applications whose workings are beyond the comprehension and control of the human user.
Ensuring data privacy, confidentiality and security.17 24 27AI developers must ensure they adhere to legal and community expectations regarding privacy, confidentiality and security of health and medical data.
Ensuring access and equity.24–26AI applications shown to be effective must be equitably accessible to low income, remote or other disadvantaged populations, and not be concentrated in already well-served populations with well-structured digital and data infrastructures.
  • AI, artificial intelligence.