Design and Implementation Issues

The early 1970s were a time of great optimism for researchers in the field of medical artificial intelligence. The initial successes of systems such as MYCIN, CASNET, and the Leeds abdominal pain system made it reasonable to assume that it was only a matter of time until computers became a standard part of physicians’ diagnostic armamentarium. Over the past few years, the emphasis in clinical decision support has shifted from its initial narrow focus on diagnostic expert systems to a much broader range of applications. Increasingly, clinicians have access to alerts, reminders, and patient-specific advice for such common tasks as prescription writing and test ordering. Despite these gains, CDSS are not yet common in patient care settings. This chapter will examine the key design and implementation concerns that must be addressed if these systems are to realize their full potential. What accounts for this lack of use? The 30-year experience described by Engle provides valuable insight into the problems encountered in the creation and deployment of CDSS. He provides a list of factors divided into critical and noncritical that he feels account for the difficulties in building a useful system and the rejection of diagnostic systems by clinicians.According to Engle,“Factors that play a role but are not critical include inadequate computers and peripheral devices, difficulty some people have working with computers, systems not user-friendly, physicians’ high regard for their own capabilities, and fear of computer competition, as well as the limited nature of the programs. In our estimation, the critical impediment to the development of decision programs useful in medicine lies in the impossibility of developing an adequate database and an effective set of decision rules.”The findings of Berner et al. help us to understand some of the frustration noted by Engle. In their test of four general diagnostic systems, Berner et al. found that “. . . the proportion of correct diagnoses ranged from 0.52–0.71 and the mean proportion of relevant diagnoses ranged from 0.19–0.37. . . .” This is hardly the type of performance which encourages use by a busy clinician. While this level of performance may be problematic for the broad-based systems like QMR, Iliad, and DXplain, programs with