From: AAAI Technical Report WS-93-01. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved. Case-Based Selection of Diagnostic Imaging Procedures Charles E. Kahn, Jr. Departmentof Radiology Medical College of Wisconsin Milwaukee, Wisconsin 53226 USA ckahn@its.mcw.edu Selecting the most appropriate diagnostic imaging procedure is a complextask: physicians must weigh the cost, performanceand risks of each procedure. ISIS (Intelligent Selection of ImagingStudies) is developmentalcase-based decision-support system that helps physicians select diagnostic imagingstudies. Thesystem encompassesboth episodic and prototypic cases in the domainof radiological proceduressuch as computedtomography(CT), ultrasound, magnetic resonance imaging (MRI)and angiography. Eachcase includes the procedurerequested, the clinical informationprovidedand the procedureactually performed.Eachcase also includes clinical questions to be asked of the referring physician to determinethe appropriateness of the requested imagingprocedureand to assist the radiologist in formulatinga diagnosis. The system contains both successful cases and "failures," such as cases with complications or where the procedure performed was deemedto be inappropriate. There are usually several (if not many)valid indications for each imagingprocedure, and there are manyfactors that can lead one to modifythe imaging protocol or to select an alternative procedure.Thesophistication of the systemwill comefromits breadth of knowledgeabout indications and contra-indications for various procedures. Thecase base has a distributed representation, and the features indexedallow the system to explain its reasoning. Indexedfeatures are organizedhierarchically to allow conceptualgeneralization. At present, the index is maintained manually. Ultimately this process will becomeautomatedor semi-automated,in which the systemwill create generalizations inductively, or will retrieve old cases and promptthe expert user to identify the differences betweenthe current case and relevant prior cases to understandthe exceptionsor fine details of the generalized knowledge.The primarygoals of the system’s knowledgeacquisition process are to incorporate the valid indications for imagingproceduresand to identify the salient clinical information necessaryto interpret them. In retrieving cases, the system gives preference to prototypic cases, becausethese cases represent the compositeexperienceof several learned cases. Wherea retrieved episodic case presents a serious conflict for the proposedplan, that case, too, will be retrieved to use in adaptationof the imagingplan. The"statistics" slot of each case contains informationto indicate the type of case (episode vs. prototype). If the case is prototype, the numberof cases whichcomprisethat case will be encodedin the statistics field to indicate the "weightingfactor" of the case’s importance. Currently, the systemincludes cases relative to only one imagingprocedure, abdominalultrasound. After validating the system’s indexing and retrieval mechanisms in this limited domain,wewill present up to 500 requests for imagingproceduresas cases to ISIS. This large numberof cases will providea sufficiently rich memoryto allow non-trivial reasoning. The system will then be ready for thorough validation of its knowledge and reasoning,and for clinical trial in comparisonto physiciansat various levels of expertise. It is this project’s ultimate goal to embedISIS within the clinical functions of our institution’s radiology information system. 164