Integrating trust into artificial intelligence for medicine: using diabetes as the exemplar disease.

Artificial Intelligence (AI) has the potential to impact healthcare across multiple domains. In diabetes, a complex chronic disease affecting 600 million people globally, AI is already being used from primary care to tertiary specialist care to reduce patient and clinician burden. However, for medical AI to be widely implemented and applied specifically to diabetes, such stakeholders as patients, clinicians, healthcare administrators, regulators, and AI developers will need to establish trust in this technology. Building trust is a balancing act depending on individual priorities of stakeholders which may not necessarily align. Both probabilistic outputs and “top-choice only” outputs are used in medical AI. To achieve trust in AI for diabetes care, it will be necessary to move beyond expecting only single, deterministic outputs and to establish clear standards for medical AI provenance and performance. This article presents priorities for each of the various stakeholders if they are to develop trust in medical AI and their responsibilities for contributing to the establishment of trust in medical AI. For a medical AI system to be trustworthy, six key attributes must be incorporated including accuracy, reproducibility, privacy/security, transparency, human oversight, and fairness. We present practical methods to achieve each of these six attributes of trustworthy medical AI prioritizing diabetes that are important for all stakeholders.