Table 1

A list of prominent issues of medical artificial intelligence (AI) applications and potential solutions

IssuePotential solution
The Frame Problem (the difficulty in identifying and updating a set of axioms to properly describe the environment for autonomous agents).
  • Clinician review of input with atypical feature statistics.

  • Rigorous prospective clinical trials in diverse patient populations.

  • Continual calibration by human feedback.

Trust in the performance of the AI program.
  • Disclosure of implementation details, nature of training sets and shortcomings of the AI systems.

  • Develop interpretable machine learning models.

  • Patient education.

Amplifying biases presented in the historical data.
  • Ongoing acquisition of training data reflecting current practice and population characteristics.

  • Identify confounders in the association-based models.

Clinical workflow disruption.
  • Redesign workflow that enables AI assistance without encouraging clinician decision-making passivity or aggravating ’alert fatigue’.