How Can Computer Programs Reason?

Computer programs are increasingly being called on to suggest or to make decisions in medical applications. Traditional methods of decision making based on flowcharts and probabilistic classification have proven to be too cumbersome to apply to large domains. As a result, programs employing artificial intelligence methods were introduced in the 1970’s. Deficiencies in those methods limited the capabilities of the resulting programs and suggested many research problems now being pursued in medical AI. Current research plans call for progress in research on representing the causal, pathophysiologic basis of disease, temporal evolution of cases, likelihood and utility, and the use of past experience in problem-solving. In addition, efforts to integrate progress in various areas and to demonstrate the applicability of the new methods are also key. A new generation of programs will result, with more sophisticated capabilities to bring to bear on the difficult problems of medical reasoning. The computing elements of health care are becoming more and more significant, and increasing sophistication is demanded of the decision-making components of medical devices. Although this trend is generic to all of medicine, it is perhaps most advanced and clearly visible in the types of instruments that are the focus of the cardiostim meeting. This paper describes developments in the application of artificial intelligence (AI) methods to building programs that reason in a manner approximating the performance of human health practitioners. In addition, it points to some of the current outstanding research problems in medical AI and suggests some promising current research directions.