The TOPOSCOUT expert system for stroke localization. 1195 The TOPOSCOUT Expert System for Stroke Localization Klaus Spitzer, PhD, MD, Andreas Thie, MD, Louis R. Caplan, MD, and Klaus Kunze, MD Clinically, strokes are localized by the findings on neurologic examination, TOPOSCOUT is an expert system designed to diagnose the anatomic location and the corresponding vascular territory of strokes based on the clinical signs and symptoms. The inference engine of TOPOSCOUT uses a backtracking algorithm and a rule-based data base that includes associations of neurologic signs with vascular and anatomic areas, TOPOSCOUT is capable of detecting typical stroke patterns, for example, "top-of-the-basilar" or Wallenberg's syndromes. The accuracy of TOPOSCOUT'S diagnoses has been tested for conformity with the final diagnoses of 129 patients in the Hamburg Stroke Data Bank, and a high level of agreement was found for hemispheric lesions. The program runs on microcomputers with MS-DOS and is intended as a practical aid for physicians not fully familiar with topologic stroke diagnosis and as an interactive teaching device. (Stroke 1989;20:1195-1201) The topologic localization of stroke is basedon detailed knowledge of the associationsof clinical signs and symptoms with lesions of particular neuroanatomic structures. Frequently, these associations can be formulated as a rule: If there is hemilateral impairment of sensation and no motor weakness or cognitive defects, then assume a lesion in the contralateral thalamus. The data are subjected to multiple pattern-matching rules to arrive at the probability of the final localization. Rule- based expert systems have been used to analyze gait abnormalities, drug interactions, headache and facial pain, and other medical problems. 1~5 We present TOPOSCOUT, a prototype microcomputer- based expert system for stroke localization, TOPO- SCOUT may be used independently or as a supple- ment to MICROSTROKE, which is an expert system for the diagnosis of stroke type.6 Materials and Methods TOPOSCOUT'S knowledge data base is arranged entirely in the form of rules, presently 171. All data are derived from anatomic charts and textbooks7-15 From the Neurologische Universitatsklinik Hamburg- Eppendorf, Hamburg, Federal Republic of Germany (K.S., A.T., K.K.) and the Department of Neurology, Tufts-New England Medical Center, Boston, Massachusetts (L.R.C.). Presented in part at the 40th Annual Meeting of the American Academy of Neurology, Cincinnati, Ohio, April 1988. Address for correspondence: Dr. Spitzer, Neurologische Uni- versitatsklinik Hamburg-Eppendorf, MartinistraBe 52, D-2000 Hamburg, Federal Republic of Germany. Received June 21, 1988; accepted March 9, 1989. and from 63 clinical publications. Rules are classi- fied into six groups: 1. Signs/symptoms rules (n=42; an example is Rule S37 in Figure 1). TOPOSCOUT is able to find a syndrome by logical deduction from clinical signs provided by the physician-user. Thus, formulation of more complicated diagnostic rules using complex terms substituted for signs is feasible. Questions to the physician-user are avoided if their answers can be logically deduced. 2. Hemisphere/brainstem rules (n=30; an exam- ple is Rule H29 in Figure 1). To estimate the odds for a hemispheric vs. a brainstem lesion, TOPOSCOUT maintains one account for each location. A priori odds were taken from stroke textbooks. 13~1S Depend- ing on the signs in a given patient, TOPOSCOUT multiplies each account by a factor listed in associ- ated rules. Relevance is an additional multiplication factor that takes into account the relative weight of a clinical item in the diagnostic decision. 3. Right/left rules («=21; an example is Rule R14 in Figure 1). These rules comprise logical deduc- tions for detection of the side of a stroke. Right- and left-sided strokes can be detected in the same patient; thus, discovery of bilateral hemispheric or brainstem lesions is possible. 4. Vascular pattern rules (n=26; an example is Rule V08 in Figure 1). Vascular patterns are coded into logical rules. As for the laterality of a stroke, conclusions from several rules can be considered at the same time, corresponding to the involvement of more than one vascular territory. At present, TOPO- SCOUT'S knowledge data base includes only patterns D ow nloaded from http://ahajournals.org by on A pril 5, 2021 1196 Stroke Vol 20, No 9, September 1989 RULE S3 7: IF AID AID AID THEI thsre is weakness in light arm there is no weakness in left arm there is weakness in right lag there is no weakness in left lag there is right heaiparesis; RULE H29 IF thara is dysarthria OB there is impairment of lower cranial narvas TKEI relevance = 2, multiply hemispheric account with 0.3, multiply brainstea account with 2.0, RULE R14: IF the patiant omits tha laft side of a drawn figure AID ha displays abnormal angles and proportions in a drawn figure THEI assuae lesion on tha right (hemispheric) sida; RULE V08: IF thare are no reliable signs of a lesion in tha territory of tha middle cerabral artary AID thare is dyslaxia AID thara is no dysgraphia THEI assuae a lesion in the territory of the posterior cerebral artary; RULE A20 IF thara is aydriasis on tha left AID there is paresis of laft aye auscles AID thare is ptosis on tha laft AID thara is right facial weakness AID thara is right heaiparesis THEI assume WEBER's syndrome on the l e f t , RULE HO9 IF tha c h o l a s t a r o l l e v e l i s a l e v a t a d THEI r e l e v a n c e = 1, m u l t i p l y HCA account w i th 1 . 3 ; m u l t i p l y ICA account w i th S 3 , FIGURE 1. Excerpt from TOPOSCOUT'S knowledge data base. for the main stems of the hemispheric cerebral arteries, the anterior cerebral artery (ACA), the middle cerebral artery (MCA), and the posterior cerebral artery (PCA). Although in 90% of patients the PCA arises from the basilar artery,13 the PCA is analyzed with the ACA and MCA as a cortical supply vessel. So far, TOPOSCOUT cannot distinguish segments and branches of these main arteries. The knowledge data base includes rules that take into account the age-dependent frequency distribution of vascular territories involved in stroke samples.16 5. Anatomic pattern rules («=28; an example is Rule A20 in Figure 1). These rules comprise pat- terns of neurologic signs associated with particular anatomic areas. For example, TOPOSCOUT can detect thalamic stroke and specific brainstem lesions such as Wallenberg's syndrome. We have included rules that provide a tolerance in diagnosis in case a single sign usually associated with a particular syndrome is missing, but the presence of other signs yields a high probability for this syndrome. 6. MCA/ICA rules (n=24; an example is Rule M09 in Figure 1). If stroke in the territory of the MCA is assumed and if there are no signs of an embolic source in the heart or proximal vascular system, a particular set of rules is activated, TOPO- SCOUT tries to assign the vascular lesion to either the internal carotid artery (ICA) or the MCA; data were derived from an investigation of occlusive disease of the MCA by Caplan et al.17 Similar to the estimation of hemispheric vs. brainstem lesion, accounts for both vessels are maintained, resulting in final odds for an ICA vs. an MCA lesion. TOPOSCOUT'S inference engine is based on a back- tracking algorithm and is the section of the com- puter program that includes the specific problem- solving capabilities of the expert system. A backtracking mechanism18-20 is a specific method of problem solving in which the expert system starts with the final goal (G). For TOPOSCOUT, G is to detect stroke laterality, hemispheric vs. brainstem location, and vascular and anatomic patterns. The inference engine searches the data base for rules for which the conclusion is R1-»G. To establish these rules, the inference engine looks for other rules that yield the premises of the first ones (R2->R1). In this D ow nloaded from http://ahajournals.org by on A pril 5, 2021 R/L deteotlon ACA assumed Spitzer et al TOPOSCOUT Expert System 1197 basic I examination! hsm/brtt dataotlon ham lesion assumed brtt latlon aaaumad vase tarr detaotlon oattarn detaotlon R/L detaotlon I pattarn Ideteotlon I MCA/ICA assumed PCA assumed MCA/ICA dateotlon FIGURE 2. Flowchart of TOPOSCOUT'sprogram architecture, hem, hemispheric; brst, brainstem; R/L, right vs. left; vase terr, vascular territory; ACA, anterior cerebral artery; MCA, middle cerebral artery; ICA, internal carotid artery; PCA, posterior cerebral artery; thromb, thrombosis. way, TOPOSCOUT backtracks to rules the premises of which are facts (F) about the current patient obtained by the physician-user. G is then concluded by chaining the conclusions F-». . . R2-»R1-»G. The inference engine includes 52 rules strictly separated from the knowledge data base. As described above, the heuristic character of TOPOSCOUT allows none, one, or more than one solution for the problems right/left, vascular pattern, and anatomic pattern. Mathematical calculations are restricted to the solu- tion of the problems hemispheric vs. brainstem and MCA vs. ICA lesion. The inference engine has access to an ASCII-text file that contains informa- tion used for explanatory and educational purposes. TOPOSCOUT asks the physician-user interactively for details of the neurologic examination, usually in a multiple-choice format. Some questions are omit- ted if the answer can be deduced logically. We included clinical items from the questionnaire of Ropper et al21 for computer-assisted data acquisi- tion in a neurologic intensive care unit, TOPOSCOUT does not use laboratory findings. TOPOSCOUT is linked to a data base in which all personal cases, including the physician-user's final diagnoses confirmed by laboratory studies, are stored. To assess the validity of TOPOSCOUT'S diag- noses, control programs calculating success rates have been implemented. In addition, special sub- routines are capable of retrieving data on authentic cases from large stroke registries, passing them to TOPOSCOUT, and driving the expert system in an automatic fashion. TOPOSCOUT'S programs are written with the VP-Expert22 expert system development shell. An IBM-compatible microcomputer with a floppy disk drive is required. Results Operating TOPOSCOUT is easy; no knowledge of computers is required. Figure 2 shows a flowchart of the program architecture, TOPOSCOUT starts with acquisition of data about a patient's neurologic signs (Figure 3). The physician-user selects one of the options presented in a multiple-choice format. Numerical input is required for some questions such as age. A summary of the patient's findings is displayed and simultaneously written to a disk file for later review. Initially, the system estimates odds for stroke localization in the cerebral hemispheres vs. in the brainstem. In both instances, TOPOSCOUT tries to determine the side of the stroke. For hemispheric lesions, TOPOSCOUT looks for patterns of signs asso- ciated with territories of the main cerebral arteries. If involvement of the MCA is assumed, TOPOSCOUT attempts to distinguish between a vascular lesion in the ICA or the MCA. For this purpose, historic and clinical data about the patient is requested. In addition, the patient's data profile is reviewed for patterns of signs associated with specific anatomic areas. If the presence of a particular syndrome is suspected and TOPOSCOUT needs more information to prove or refute it, the physician-user is consulted again. It takes approximately 3-4 minutes to enter data for a patient. At the end of the session, TOPOSCOUT has summed up all entered information to present its final diag- nosis (Figure 4), which comprises odds for hemi- spheric vs. brainstem lesions and TOPOSCOUT'S pre- diction including the supposed side of the stroke, an anatomic pattern if present, and TOPOSCOUT'S esti- mation of the involvement of large arteries, com- bined with their age-matched odds. D ow nloaded from http://ahajournals.org by on A pril 5, 2021 1198 Stroke Vol 20, No 9, September 1989 How about f a c i a l weakness 7 no weakness L weakness < b i l a t e r a l weakness R weakness How about f a c i a l sensation ? normal < L impairment b i l a t e r a l impairment How about the lower cranial nerves ? no d e f i c i t < L impairment b i l a t e r a l impairment Do nystagmus or vertigo e x i s t ? no < yes Is there dysarthria ? no < yes E impairment R impairment FIGURE 3. Excerpt from sample session: TOPOSCOUT'S data acquisition. L, left; R, right. Enter to select EID to conplete /Q to Quit ' for Unknown Before leaving TOPOSCOUT, the physician-user is prompted to enter the clinical diagnosis based on results of laboratory studies, if available. This authen- tic information is also stored in the disk file. TOPOSCOUT is able to explain its conclusions. Comments about particular keywords include infor- mation about the pertinent literature taken into account at different phases of diagnosis. To test the quality of TOPOSCOUT'S diagnoses, we have transferred data on 129 patients from the Hamburg Stroke Data Bank23 to TOPOSCOUT'S con- trol unit, TOPOSCOUT'S diagnoses were compared with the final diagnoses in the stroke registry based on the results of laboratory examinations, TOPO- SCOUT recognized 87% of all hemispheric strokes regardless of side (76% of right hemispheric strokes and 74% of left hemispheric strokes) and 56% of all brainstem strokes regardless of side (56% of right brainstem strokes and 50% of left brainstem strokes). TOPOSCOUT identified 86% of all MCA-territory strokes regardless of side (76% of right MCA strokes and 73% of left MCA strokes) and 41% of all PCA-territory strokes regardless of side (50% of right PCA strokes and 36% of left PCA strokes). There were no ACA-territory strokes. Discussion TOPOSCOUT is a microcomputer-based expert sys- tem designed to assist physicians in diagnosis of the topology of stroke using information entirely avail- able at the bedside. Simulating the human diagnos- tic approach, TOPOSCOUT'S pattern-matching algo- rithms can direct the physician-user's attention to syndromes that may occasionally be overlooked. TOPOSCOUT'S topologic knowledge data base repre- sents a fast and easily accessible reference for physicians involved in the management of stroke patients. Like many other expert systems in medicine,24 TOPOSCOUT serves as a tool for clinical teaching and may be useful for planning further diagnostic procedures. In addition, since all data of personal cases, including the final diagnoses con- firmed by laboratory studies, are stored, TOPOSCOUT provides a stroke registry. Since the computer file encompasses the com- plete neurologic status of a patient, the physician- user wastes less time with handwritten charts. Tak- ing into account the fast retrieval of the personal stroke registry, we believe that the few minutes spent for data entry are well invested and do not slow down the physician-user in the setting of a busy practice. TOPOSCOUT'S inference engine has been devel- oped applying artificial intelligence techniques in a medical diagnostic process.25-26 The inference engine comprises goal-directed programming, pattern matching abilities, and explanatory capabilities. These methods must be distinguished from decision-support programs based on statistical pro- cedures.27-29 We consider statistically based meth- Odds for hemispheric lesion: 88.3 Odds for brainstem lesion 11 7 Predicting lesion in right hemisphere. Pattern: Pure motor stroke (basis pontis/internal capsule). Predicting lesion in AC* t e r r i t o r y . Predicting lesion in HCA territory: yes Predicting lesion in PCA t e r r i t o r y (age-matched odds: 6X) (age-matched odds: 8lX> (age-matched odds 13X) FIGURE 4. Excerpt from sample session: TOPOSCOUT'S diagnosis. ACA, anterior cere- bral artery; MCA, middle cerebral artery; PCA, posterior cerebral artery. D ow nloaded from http://ahajournals.org by on A pril 5, 2021 Spitzer et al TOPOSCOUT Expert System 1199 ods, especially Bayesian techniques,30-31 to be inad- equate tools for a topologic diagnosis-support system; the set of topologic hypotheses is generally not exhaustive and mutually exclusive as assumed by Bayesian statistics.29-32 Thus, the frequent involvement of two or more cerebral territories by a stroke contradicts the assumption of only one cor- rect diagnosis in each case as required by most statistical procedures. In TOPOSCOUT, mathematical calculation is limited to a few minor problems in which general objections to statistical procedures do not have to be considered. Adversely, the MICRO- STROKE expert system for the diagnosis of stroke type uses modified Bayesian inference techniques.6 The backtracking algorithm was selected to restrict data acquisition to those questions that are needed to pursue relevant goals, TOPOSCOUT skips queries if pertinent information is supplied by other answers or if questions seem to be of limited value in the current diagnostic process; for example, not all neuropsychological phenomena33 are requested for each patient. However, the inference engine simu- lates forward-chaining features by requiring the physician-user to enter complete results of a neuro- logic examination, providing the stroke registry capability. In this way, later reassessment of TOPO- SCOUT'S decision-making process is feasible. Published programs for computer-assisted topo- logic diagnosis in neurology can be classified into two groups based on their neuroanatomic expertise.34 In one group, knowledge data bases interpret the spatial representation of neuroana- tomic items and their relations.34-39 In the second group, knowledge data bases are constructed by associations of neuroanatomic items with clinical signs without considering the structural or func- tional relations of anatomy. Localization is pre- dicted by recognizing patterns of clinical findings.40-43 TOPOSCOUT belongs to this second group. The lack of total knowledge of neural con- nections in the central nervous system precludes storage of all structural and functional relations in a data base. Pattern recognition is an appropriate approach for this representation of rudimentary knowledge, and pattern matching provides the capa- bility to detect lesions at multiple locations. How- ever, TOPOSCOUT shares the disadvantage of all expert systems based on associations of neuroana- tomic items with clinical signs: it is limited to recognizing common clinical patterns and is unable to diagnose patients whose syndromes have not been previously described or have not been included in the rule-based system. At present, there are still other drawbacks to our expert system. One serious handicap is that validity control has shown poor detection of brainstem lesions and involvement of the PCA. Too few rules (31) have been implemented concerning topologic diagnosis of brainstem lesions (compared with 72 rules for hemispheric lesions). If a symptom could be caused by a hemispheric or a brainstem lesion, the former location is often falsely favored by TOPOSCOUT because of probability assignment and disregard of pertinent cosymptoms. Infra ten torial topologic diagnosis forms an obstacle not only for TOPOSCOUT but for other systems as well. For exam- ple, PAL was correct in detecting posterior fossa strokes in only 53% of cases.42-43 Posterior circula- tion stroke is more heterogeneous than anterior circulation stroke. A larger registry of patients with vertebrobasilar stroke will be needed to amplify TOPOSCOUT'S rule system. Most diagnostic failures were caused by TOPO- SCOUT'S insufficient tolerance. Stroke location was often not correctly detected because an appropriate minor symptom was missing, prompting the definite rejection of this location despite other evidence. Some rules adhere too closely to logical links of clinical and anatomic items, a fact that uncovers a general disadvantage of rule-based systems, TOPO- SCOUT was helpless in identifying lesions associated with a broad spectrum of clinical symptoms, such as caudate infarcts.44 We must improve our knowl- edge data base to include more rules that take into account variability of clinical syndromes. Furthermore, the knowledge data base should be expanded to include clinical patterns of stroke in cerebral arteries supplying brainstem structures and patterns associated with stroke in terminal parts of hemispheric arteries. Lacunar infarcts should be identified by TOPOSCOUT. Additional improvement will be achieved by using TOPOSCOUT'S self-learning characteristics when data from past consultations are included in current diagnostic decisions. The rules implemented in this prototype expert system should be regarded as preliminary. They are still inadequate concerning consistency, complete- ness, and independence. We will attempt to over- come these shortcomings by applying knowledge acquisition and verification tools.3'45-46 Since 1979, numerous expert systems have been successfully developed with expert system shells (software tools for developing expert systems).47-49 In contrast, such a shell seems to be too inflexible for this specific stroke localization problem. In particular, inclusion of tolerance in the mechanism of pattern recognition seems to pose severe prob- lems for a shell. To provide TOPOSCOUT with full deductive reasoning capabilities, the expert system is translated into PROLOG.50 In the future, a prospec- tive study will be carried out to verify the validity of the program. This prototype expert system suggests that computer-assisted stroke localization is feasible at the bedside and that, with further improvement, TOPOSCOUT holds promise as a tool in the clinical routine of stroke diagnosis. Finally, we clearly do not deny that skillful and experienced neurologists, with their present 80-86% accuracy rate,51 have legitimate reasons to criticize the competence of the newborn TOPOSCOUT. But we also want to empha- size that these authorities decided for good reasons D ow nloaded from http://ahajournals.org by on A pril 5, 2021 1200 Stroke Vol 20, No 9, September 1989 to spend a considerable number of years as medical students, residents and, possibly, stroke fellows. Therefore, we believe that it is tolerable to accept a training period of several years for the development of an expert system's software and expansion of its knowledge data base before it is compared with the most competent in the field. References 1. Davis R, Buchanan B, Shortliffe E: Production rules as a representation for a knowledge based consultation program. Artif Intelligence 1977;7:15-45 2. Dzierzanowski JM, Bourne JR, Shiavi R, Sandell HS, Guy D: GAITSPERT: An expert system for the evaluation of abnor- mal human locomotion arising from stroke. IEEE Trans Biomed Eng 1985;32:935-942 3. Hudson DL, Estrin T: Derivation of rule-based knowledge from established medical outlines. Comput Biol Med 1984; 14:3-13 4. Matsumura Y: RHINOS: A consultation system for diagnosis of headache and facial pain. Comput Methods Programs Biomed 1986;23:65-71 5. Roach J, Lee S, Wilcke J, Ehrich M: An expert system for information on pharmacology and drug interactions. Comput Biol Med 1985;15:ll-23 6. Spitzer K, Thie A, Caplan LR, Kunze K: MICROSTROKE—A microcomputer-based expert system for stroke type diagno- sis. Stroke (in press) 7. Brazis PW, Masdeu JC, Biller J: Localization in Clinical Neurology. Boston, Little, Brown & Co, 1985 8. Brodal A: Neurological Anatomy. New York, Oxford Uni- versity Press, 1981 9. Chusid J: Correlative Neuroanatomy, Functional Neurol- ogy. Los Altos, Calif, Lange, 1985 10. Duus P: Topical Diagnosis in Neurology. Stuttgart, Thieme, 1983 11. Niuewenhuys R, Voogd J, v Huijzen C: The Human Central Nervous System. A Synopsis and Atlas. Berlin, Springer, 1981 12. Poritsky R: Neuroanatomical Pathways. Philadelphia, WB Saunders Co, 1984 13. Barnett HJM, Mohr JP, Stein BM, Yatsu FM (eds): Stroke. II1I. New York, Churchill Livingstone Inc, 1986 14. Caplan LR, Stein RW: Stroke. Boston, Butterworths, 1986 15. Millikan CH, McDowell FM, Easton JD: Stroke. Philadel- phia, Lea & Febiger, 1987 16. Spitzer K, Thie A, Fuchs B, Kunze K: Ischemic strokes in young adults (abstract). J Neurol 1988;235(suppl 1):5 17. Caplan LR, Babikian V, Helgason C, Hier DB, DeWitt D, Patel D, Stein R: Occlusive disease of the middle cerebral artery. Neurology 1985;35:975-982 18. Hayes-Roth F, Waterman DA, Lenat DB (eds): Building Expert Systems. Reading, Mass, Addison-Wesley, 1983 19. Waterman DA: ,4 Guide to Expert Systems. Reading, Mass, Addison-Wesley, 1986 20. Nilsson NJ: Principles of Artificial Intelligence. Berlin, Springer, 1982 21. Ropper AH, Griswold K, McKenna D, Souder D: Computer- guided neurologic assessment in the neurologic intensive care unit. Heart Lung 1981;10:54-60 22. VP-Expert. Ruled-Based Expert System Development Tool. Berkeley, Calif, Paperback Software International, 1987 23. Spitzer K, Becker V, Thie A, Kunze K: The Hamburg Stroke Data Bank: Goals, design and preliminary results. J Neurol 1989;236:139-144 24. Szolovits P: Using artificial intelligence models of medical decision-making in medical education, in Pages JC, Levy AH, Gremy F, Anderson J (eds): Meeting the Challenge: Informatics and Medical Education. New York, Elsevier, 1983, pp 271-281 25. Szolovits P: Artificial Intelligence in Medicine. Boulder, Colo, Westview Press, 1982 26. Torasso P: Knowledge based expert systems for medical diagnosis. Stat Med 1985;4:317-325 27. Fox J, Barber D, Bardhan KD: Alternatives to Bayes?—A quantitative comparison with rule-based diagnostic infer- ence. Methods Inf Med 1980;19:210-215 28. Schwartz B, Patil RS, Szolovits P: Artificial intelligence in medicine. Where do we stand? N Engl J Med 1987; 316:685-688 29. Bouckaert A: Medical diagnosis: Are expert systems needed? IntJ Biomed Comput 1987;20:123-133 30. Salamon R, Bernadet M, Samson M, Derouesne C, Gremy F: Bayesian method applied to decision making in neurology— Methodological considerations. Methods Inf Med 1976; 15:174-179 31. Duda RO, Hart PE, Nilsson NJ: Subjective Bayesian Meth- ods for Rule-Based Inference Systems. Stanford, Calif, National Computer Conference, 1976, pp 1075-1082 32. Szolovits P, Pauker SG: Categorical and probabilistic rea- soning in medical diagnosis. Artif Intell 1978;11:115-144 33. Pinskaya LB: Neuropsychological phenomena of secondary stem dysfunction in patients with strokes of hemispheric localization. Neurosci Behav Physiol 1986;16:101-104 34. First MB, Weimer BJ, McLinden S, Miller RA: LOCALIZE: Computer-assisted localization of peripheral nervous system lesions. Comput Biomed Res 1982;15:525-543 35. Meyer AU, Weissman WK: Computer analysis of the clini- cal neurological examination. Comput Biol Med 1973; 3:111-117 36. Meyer AU, Weissman WK: The localization of human nervous system lesions by computer analysis. Trans Am Neurol Assoc 1973;98:327 37. Catanzarite VA, Greenburg AG, Bremermann HJ: Com- puter consultation in neurology: Subjective and objective evaluations of the "NEUROLOGIST" system. Comput Biol Med 1982;12:343-355 38. Catanzarite VA, Greenburg AG, Bremermann HJ: Com- puter assisted diagnosis and computer consultation in neu- rology: Preliminary testing of diagnostic accuracy for the NEUROLOGIST system. Int J Neurosci 1981;13:43-54 39. Banks G, Weimer B: Symbolic coordinate anatomy for neurology (SCAN). J Med Syst 1984;8:157-162 40. Reggia JA: A production rule system for neurological local- ization, in Reggia JA, Tuhrim S (eds): Computer-Assisted Medical Decision Making. II. New York, Springer, 1985, pp 49-62 41. Stewart A, Cala L: Mathematical method to utilize a com- puter for diagnosis of site and type of intracerebral mass lesions. BrJ Radiol 1975;48:97-100 42. Hier DB, Caplan LR, Hill H, Evans M, Sinha A: A microcomputer-based expert system to assist in localization of anatomic damage after stroke (abstract). Neurology 1984; 34(suppl 1):83 43. Hier DB, Atkinson GD, Perline R, Hill H, Evans M, Desai B, McCormick WC, Caplan LR: Can a patient data base help build a stroke diagnostic expert system? Med Inf 1986; 11:75-81 44. Caplan LR, Schahmann JD, Baquis GD, Kase CS, Feld- mann E, Greenberg JP, Gorelick PB, Helgason C, Hier DB: Caudate infarcts (abstract). Neurology 1988;38(suppl 1):262 45. Suwa M, Scott AC, Shortliffe EH: An approach to verifying completeness and consistency in a rule-based expert system. Artif Intell Mag 1982;2:16-21 46. Mars NJ, Miller PL: Knowledge acquisition and verification tools for medical expert systems. Med Decis Making 1987; 7:6-11 47. van Melle W: A domain-independent production rule system for consultation programs, in Proceedings of the Sixth International Joint Conference on Artificial Intelligence. Tokyo, 1979. Los Altos, Calif, Morgan Kaufmann, 1979, pp 923-925 48. Weiss S, Kulikowski C: EXPERT: A system for developing consultation models, in Proceedings of the Sixth Interna- tional Joint Conference on Artificial Intelligence. Tokyo, 1979. Los Altos, Calif, Morgan Kaufmann, 1979, pp 942-960 D ow nloaded from http://ahajournals.org by on A pril 5, 2021 Spitzer et al TOPOSCOUT Expert System 1201 49. Blomberg DJ, Guth JL, Fattu JM, Patrick EA: Evaluation of a new classification system for anemias using Consult Learn- ing System. Comput Methods Programs Biomed 1986; 22:119-125 50. Townsend C: Mastering Expert Systems With Turbo Prolog. Indianapolis, Sams, 1986 51. Nadeau SE, Jordan JE, Mishra SK: A prospective evalua- tion of the localizing value of clinical data in acute stroke (abstract). Neurology 1988;38(suppl 1):344 KEY WORDS • cerebrovascular disorders • expert systems D ow nloaded from http://ahajournals.org by on A pril 5, 2021