Plan Recognition and Revision in Support of Guideline-Based Care

Weconsider the problem of providing automatedsupport for guideline-basedclinical care. Clinical guidelines are a common formatin medicaldomainsfor prescribinga set of rules andpolicies that an attendingphysicianshouldfollow. In terms of an AI planningtask, clinical guidelinescan be viewedas a sharedlibrary of highlyreusableskeletal reactive plans, whosedetails need to be refined by the executingplanner over significant periods of time. Theapplication of clinical guidelines involves the collection and interpretation of patient-related data, the application of prespecified plans, and a revision of the plans when necessary. Overthe past decade, several research groups haveimplemented reactive-planningarchitectures specific for the task of refining skeletal plans over time. Theimportance of such systemsis increasing as moreclinical data are being capturedand represented in an electronic format, and as quality control of medical care growsin importance. Conductinga flexible, intelligent dialoguewith the physicianuser of these systemsrequires an ability to reason about the user's goals and plans and about possible modifications to these plans. Wepoint out that automatedsupport for guideline-basedcare can be viewedas a collaborative effort of two planning agents: the physician and an automated ("assistant") planner, whoseknowledgemight limited, or whomightnot have access to all the data. Wedemon strate that achieving even a modicum of collaboration between the two planners and of highly desirable flexibility in the automated support to the physician involves a form of reasoning about mentalstates: a recognition of eachplanner's (in particular, the physician's) intentions and plans to achievethem, anda consideration of the available plan-revision strategies during executiontime. In particular, automatedsupport for clinical guidelines could be enhanced considerably by a sharable, explicit, formal representation of (1) therapy-planning-operators' effects, (2) plan-revision strategies, and (3) the underlying goals andpolicies of the guideline, in the form of temporal-abstractionpatterns to be maintained, achieved, or avoided.Finally, relying on our analysis, wecan list in a structured manner several possible sources of disagreement between the two planners, thus supportingthe maintenance of the automated planner's knowledgebase as well as improving the appropriateness, and thus acceptability (by the physician) of its future recommendations.