COMPUTER-BASED EXPLANATION METHODS FOR DECISION MODELS

点击翻译按钮获取中文摘要

There is growing recognition that, while practice guidelines can predispose physicians to behavioral change, even the most respected and clinically relevant guidelines often do not affect actual practice. Despite the apparent reluctance of physicians to incorporate published practice guidelines into their clinical routine, there are signs that immediate automated reminders can have a profound influence on physician behavior, and on the cost and quality of medical care. As practice guidelines proliferate and become more complex, the dilemma facing those who design practice guidelines is the tradeoff between the comprehensiveness of the guidelines and the likelihood they will be retained and implemented by the physicians for whom they are intended. This same complexity can complicate computer-based implementation of guidelines. For example, the dependence of guidelines on patient preferences and individual risk factors makes the rule-based approach impractical for many guidelines. Intelligent decision systems (IDS) are specifically designed to implement complex, flexible guidelines tailored to varying clinical circumstances. These systems provide advice based on a tailored decision-analytic model. But these systems assume that a physician-user will be persuaded by a review of the decision model and its associated quantitative results. The goal of this work is to develop and test improved explanation methods for quantitative decision models so that intelligent decision systems can be used in a consultative mode. We have developed a program called QxQ, which uses symbolic reasoning to provide qualitative text explanations for the results of decision trees. This work is intended to extract the key architecture- and domain-independent elements of QxQ's methodology, and to reimplement and test them as part of MIDAS, an existing IDS. First, we will develop a domain- and system-independent version of QxQ, called EQxQ (for essential QxQ), by applying it to decision models constructed in MIDAS. In collaboration with the developers of MIDAS, we will identify the key independent elements necessary to generate explanations from an IDS. Second, we will devise and implement additional explanation methods in EQxQ to allow explanation of complex modeling constructs, such as Markov models, cycle trees, and cost-effectiveness models. Third, we will test the robustness of the MIDAS-EQxQ system by using it to implement a decision model underlying a clinically relevant practice guideline. Finally, we will establish and pilot test the routine use of the program in an outpatient clinic. We will survey the physician-users to determine the strengths and weaknesses of the program. Using pilot data from the program's routine use, we will design a prospective clinical evaluation of the system and its explanations by measuring their effect on patient outcome. Our study will provide data regarding the clinical utility of intelligent decision systems.