Chess lessons: Harnessing collective human intelligence and imitation learning to support clinical decisions

Over the past 20 years, chess-playing artificial intelligence (AI) systems have far surpassed human performance. One reason for their success is that the rules of chess are well defined, thereby allowing AI-based chess systems to learn easily which strategies optimize outcomes.

In contrast, AI systems in health care have been considerably less successful. A primary challenge to developing AI clinical decision support systems (CDSSs) that offer meaningful guidance is the uncertainty inherent in the practice of medicine. Many syndromes cannot be definitively diagnosed and even expert clinicians frequently disagree. This uncertainty also propagates into treatment decisions that are often predicated on a correct diagnosis. In contrast to chess, the rules of human biology are incompletely understood, limiting the ability of AI systems to learn ‘‘correct’’ diagnostic and treatment strategies. Thus, advanced AI approaches such as reinforcement learning, though successful in chess and other strategic games, may not be suitable for learning uncertain clinical patterns from electronic health record (EHR) data.