• information technology
• decision support, computerized
• patient safety
• implementation science
• general practice

On a clear January morning in Florida, a Tesla enthusiast and network entrepreneur was driving his new Tesla Model S on US Highway 27A, returning from a family trip. He had posted dozens of widely circulated YouTube tutorial videos on his vehicle and clearly understood many of the technical details of his car. That day, he let the vehicle run autonomously on Autopilot mode for 37 min, before it crashed into the trailer of a truck turning left. The Autopilot did not identify the white side of the trailer as a potential hazard, and the driver was killed, leaving his family and his high-tech business behind.1 This tragedy is not a metaphor for artificial intelligence (AI) applications but an example of a long-recognised challenge in AI: the Frame Problem.2 Although rarely appreciated in the scholarly and lay descriptions of the stunning recent successes of AI in medical applications, the Frame Problem and related AI challenges will have unintended harmful effects to the care of patients if not directly addressed.

With the recent advancement in machine learning algorithms, many medical tasks previously thought to require human expertise have been replicated by AI systems at or above the level of accuracy in human experts. These important demonstrations range from evaluating fundus retinography3 and histopathology4 to reading chest radiographs5 and assessment of skin lesions.6 These studies have encompassed very large numbers of patient cases and have been extensively benchmarked against clinicians. However, all these studies are retrospective in that they involve a collection of labelled cases against which the AI systems are trained and another collection against which they are tested or validated. So far, they have not entered into routine prospective use in the clinic where the Frame Problem will manifest itself most pathologically.

The Frame …