The MICROSTROKE expert system for stroke type diagnosis. 1353 The MICROSTROKE Expert System for Stroke Type Diagnosis Klaus Spitzer, PhD, MD, Andreas Thie, MD, Louis R. Caplan, MD, and Klaus Kunze, MD MICROSTROKE is a prototype expert system designed to categorize and diagnose stroke types based on clinical information. The knowledge base of MICROSTROKE includes information from large stroke registries. The system first queries the physician-user for details of the patient's history, information about the onset of stroke, accompanying symptoms, and pertinent neurologic findings and then sums the individual data items, factors in the a priori odds, and arrives at the probabilities of different stroke types for a given patient. Specific diagnosis of stroke type includes thrombosis, embolus, lacune, intracerebral hemorrhage, and subarachnoid hemorrhage. Stroke type diagnoses by MICROSTROKE were correct in 72.8% of 250 cases in the Hamburg Stroke Data Bank, MICROSTROKE runs on any MS-DOS microcomputer and is intended as a practical aid for physicians not fully familiar with the diagnosis of stroke types. (Stroke 1989;20:1353-1356) For the bedside assessment of stroke type,knowledge of the frequency distributions ofsigns, symptoms, and ecological data asso- ciated with the different stroke types can be of prime importance. Neurologists collect historical data, neurologic signs, and symptoms to arrive at a "best guess" as to stroke type, which then forms the basis for performing further diagnostic proce- dures such as computed tomography (CT scan) or cardiologic or cerebrovascular tests.1 During the last decade, much research effort has been devoted to the development of expert systems to cope with complex medical decision-making. An expert system is "an embodiment within a com- puter of a knowledge-based component, from an expert skill, in such a form that the system can offer intelligent advice or make an intelligent decision about a processing function."2 Such a system uses expert knowledge to attain high levels of perfor- mance in a narrow problem area.3 We present MICROSTROKE, the prototype of an expert system for computer-supported stroke type diagnosis based only on clinical and historical patient 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 13th International Joint Conference on Stroke and Cerebral Circulation, February 18-20, 1988, San Diego, California. Address for correspondence: Dr. Spitzer, Neurologische Uni- versitatsklinik Hamburg-Eppendorf, MartinistraBe 52, D-2000 Hamburg, Federal Republic of Germany. Received August 4, 1988; accepted May 16, 1989. data available at the bedside, MICROSTROKE serves as an aid in the diagnostic work-up of stroke as both a stroke patient data bank and as an educational tool in clinical teaching. Another expert system, TOPOSCOUT, has been independently developed for stroke localization.4 Materials and Methods MICROSTROKE includes three knowledge data bases. The first contains frequency distributions of clinical and ecological parameters for the stroke types throm- bosis, embolus, lacune, intracerebral hemorrhage (ICH), and subarachnoid hemorrhage (SAH). Data are derived from the Stroke Data Bank,5 the Michael Reese Stroke Registry,6 and the Harvard Coopera- tive Stroke Registry7 (Figure 1). The frequency distribution tables contain relative frequencies of single items (symptoms or historical data) or of a set of alternative symptoms for all stroke patients and for each stroke type, respectively. A second knowledge data base is implemented using rule-based information coding. Rule-based systems depend on the hypotheses that expert knowl- edge consists of many independent, situation- specific rules and that computers can simulate expert reasoning by stringing these rules together in chains of deduction.8 The "if" part of a rule (the premise) contains the pattern or attributes that must be matched for the rule to be used. The "then" part (the conclusion) contains an assertion to be made when the premise is satisfied. A typical rule is "If there is hypertension at onset and early course of deficit is gradual smooth progression of symptoms, D ow nloaded from http://ahajournals.org by on A pril 5, 2021 1354 Stroke Vol 20, No 10, October 1989 a l l thr l i e u b ICH SiH OnMt on i r i i i a g . 31 40 BO 1 7 1 3 I E Physical S t r t u at o n u t . 4 1 1 S 10 16 Ordinary daily a c t i v i t y a t o m t t . B9 64 47 68 64 64 Othar/oakson. 7 6 2 10 13 6 Sourer Harvard Straki R t g i i t r y . ( p a r c u t a g t ) FIGURE 1. Representation of MICROSTROKE'S knowledge base of frequency distributions, all, all stroke types; thr, thrombosis; lac, lacune; emb, embolus; ICH, intra- cerebral hemorrhage; SAH, subarachnoid hemorrhage. then display warning for ICH.'' Certain rules include combinations of symptoms that are associated with a high probability of intracranial hemorrhage. At present, MICROSTROKE matches the premises of 11 rules with data from a current patient to prompt a warning for ICH, SAH, or both if they apply. The third knowledge data base consists of exclu- sively text information, used by the tutorial module of MICROSTROKE and stored as an American Stan- dard Code for Information Interchange (ASCII) file. This knowledge data base serves only educational purposes and has no influence on calculations of stroke type diagnostic probabilities. MICROSTROKE acquires knowledge by interac- tively asking the physician, user for details of the patient's history, information about the onset of stroke, and accompanying symptoms in a question- naire comprising 38 items for which frequency distribution tables in different types of stroke are available. The answers accepted are yes, no, unknown, or an option if a multiple-choice question is presented. The inference engine of any expert system is the computer program that provides its general problem- solving capabilities. The inference engine is sepa- rated from the collection of domain knowledge, the knowledge data base, MICROSTROKE'S inference engine calculates probabilities of different stroke types using modified Bayesian inference tech- niques. Each stroke type is attributed an account. Starting from initial accounts representing the a priori odds for the five stroke types, accounts are recalculated after each question depending on the physician-user's answer. The order of the questions presented depends on the MICROSTROKE mode selected. If there is no laboratory data available, intracranial hemorrhage cannot be excluded, and MICROSTROKE'S first goal is to detect signs of ICH or SAH. If intracranial hemorrhage has already been excluded, for example, by CT scan and lumbar puncture, data are acquired to differentiate isch- emic stroke types. The accounts are displayed as probabilities of stroke types by multiplying each with a constant, yielding an account sum of 100. Figure 2 shows the calculation process of MICRO- STROKE'S inference engine in some detail. It must be emphasized that the weight attributed to an item depends on the physician-user's answer regarding this item. Thus, a positive answer may influence Let type be an element of {thrombus,embolus,lacune, ICH,SAH}, a(iype)[n] the account of type after the n-th question, and p(iype)[n] the correspondent prob- ability of type after the n-th question. The five accounts a(thrombus)[0]=34, a(embolus)[0]=31, a(lacune)[0]=19, a(ICB)[0]=W, a(SAH)[0j=B represent the a priori odds before any patient infor- mation is available. The formula "(tVP')[" + 1] = <"(1 1 Haxiial dtfieit at oistt. 2 Stuttering or itapwis* coirs*. 3 Gradial nooth C O U T H . i Flictiant coirs*. 6 Early conn* unknown. (1-5) FIGURE 3. Excerpt from typical MICROSTROKE session. ICH, intracerebral hemorrhage; SAH, subarachnoid hemorrhage. ways of reasoning, physician-user's instructions, and help screens. In the third, expert system, module, the physician- user is first asked whether intracranial bleeding has already been excluded. If so, MICROSTROKE omits questions especially designed to detect ICH and SAH. MICROSTROKE starts out with the a priori odds of different stroke types and modifies them accord- ing to the physician-user's consecutive answers. Figure 3 shows a typical MICROSTROKE screen. The upper part displays the calculated odds of different stroke types for the patient whose entered charac- teristics are listed in the middle, MICROSTROKE'S questions and the answers that will be accepted by the inference engine appear in the lower part. In this particular case, a warning message appears on the screen, drawing attention to the possibility of intra- cranial bleeding in this young patient presenting with headache. The patient's data profile is reviewed continuously during the session for patterns of symptoms commensurate with ICH or SAH. At the end of the session, MICROSTROKE has summed up all entered information to present its final odds. The physician-user is asked for a final diagnosis based on laboratory studies, if available. When all data of authentic cases are stored, including the physician-user's final diagnosis con- firmed by laboratory studies, MICROSTROKE updates its stroke registry (fourth module). Data profiles can be replayed for later reviews of particular cases, and information can also be transferred to other programs for statistical analyses or scientific reports. MICROSTROKE'S registry has the ability to store large quantities of data, providing easy access for patient care and research. MICROSTROKE'S fifth, control unit, module has access to data files of large stroke registries that are not included in its knowledge base. Cases stored in these stroke registries may be presented to MICRO- STROKE in an automatic fashion so that the data of a particular stroke registry, rather than the physician- user, will answer the system's questions. The final odds calculated by MICROSTROKE may then be com- pared with the diagnosed stroke types in the regis- try, thereby achieving quality control of the system. Using the control unit, MICROSTROKE'S guesses have been prospectively tested for conformity with the final diagnoses of 250 cases in the Hamburg Stroke Data Bank.9 MICROSTROKE was correct in 72.8% of all, in 11 of 12 SAH (91.7%), in 17 of 27 ICH (63.0%), in 96 of 128 embolic (75.0%), and in 58 of 82 thrombotic strokes (70.7%). Nine ICH, one SAH, and 12 embolic strokes were incorrectly diagnosed as thrombosis, and one lacune was taken for an embolus. The control unit also checks the quality of MICROSTROKE'S diagnostic assessment of authentic cases from previous sessions using the built-in stroke registry. Discussion MICROSTROKE is designed to serve as a computer- based diagnostic tool and a knowledge base for stroke type diagnosis, MICROSTROKE is intended as a practical aid for practitioners and neurologists involved in stroke management on wards, in stroke units, and in emergency rooms. Our expert system is hardware-independent, allowing it to be run on a wide variety of small computers, including laptops. These computers are now available in almost all medical institutions. Application of computer- assisted methods for faster and more accurate diag- nostic routines could save time, permitting physician- users to give more time to patient care or teaching.10 MICROSTROKE'S ability to detect intracranial hemor- rhage with high sensitivity could prompt physician- users to make appropriate diagnostic tests very early during the clinical course. Simulating a physician's decision-making pro- cess, MICROSTROKE incorporates modified Bayesian statistics based on frequency distributions of symp- toms combined with information stored in rule- based knowledge data banks.11-13 In contrast to statistically oriented expert systems, a class of expert systems based on inexact reasoning has been developed by investigators in artificial intelligence, taking into account the judgmental, subjective, and nonprobabilistic characteristics of medical knowl- edge.14-16 TOPOSCOUT, an expert system for stroke localization, belongs to this class of expert systems.4 Though there are major objections to the use of Bayesian techniques in medical diagnosis.17'18 Bay- esian classification systems may be more valuable than is commonly believed for some medical problems.19-20 The major advantage of computer programs based on Bayesian statistics is their abil- ity to incorporate data from large patient registries. For example, MOS is a module that deduces stroke type by using clinical and historical data from the Michael Reese Stroke Registry.21 Like MOS, many expert systems developed in the medical field are connected to clinical data banks.22 Furthermore, MICROSTROKE exhibits self-learning characteristics, making use of its stroke registry module, so that D ow nloaded from http://ahajournals.org by on A pril 5, 2021 1356 Stroke Vol 20, No 10, October 1989 with each new patient, the amount of stored infor- mation increases. If these data are made available to the inference engine, MICROSTROKE simulates the growing clinical experience of a human expert. Since expertise is the cornerstone upon which an entire expert system is built,23 the true power of MICROSTROKE lies in its knowledge data base. Yet, the process of extracting expertise from an expert physician and transforming it into a formal structure that can be understood and used by an expert system is a long, difficult, and error-prone task.16-24 For this reason, we decided to implement only published information from large stroke registries. We are aware that important clinical items are missing from MICROSTROKE'S questionnaire, but we preferred to start out with a limited but well- documented knowledge base. This restriction of the knowledge base was mainly responsible for the 27.2% incorrect stroke type diagnoses, especially in ICH and thrombotic strokes. For example, due to the lack of discriminating questions, MICROSTROKE revealed difficulties in distinguishing patients with cerebral infarcts who presented with impaired lev- els of consciousness from patients with ICH. As work progresses, MICROSTROKE'S knowledge base will be enlarged by published cases that have been studied in greater detail to improve its performance. Nevertheless, in converting this prototype expert system into a high-performance one we may meet with unexpected difficulties.25 MICROSTROKE suggests that computer-assisted stroke type diagnosis is feasible at the bedside.26 A prospective study currently in progress at Tufts- New England Medical Center and Hamburg Uni- versity Hospital must be concluded to appraise MICROSTROKE'S true value. In the future, efforts to improve MICROSTROKE'S performance will include expansion of its knowledge bases, incorporation of its own experience contained in its stroke registry, and further problem-specific modification of the underlying inference techniques. Finally, we want to emphasize that physicians disagree among themselves as much as computer programs disagree with doctors,17 especially in the diagnosis of stroke type.27 References 1. Caplan LR, Stein RW: Stroke. A Clinical Approach. Boston, Butterworths, 1986 2. Forsyth R (ed): Expert Systems. London, Chapman & Hall, 1984 3. Waterman DA: A Guide to Expert Systems. Reading, Mass, Addison-Wesley, 1986 4. Spitzer K, Thie A, Caplan LR, Kunze K: The TOPOSCOUT expert system for stroke localization. Stroke 1989; 20:1195-1201 5. Foulkes MA, Wolf PA, Price TR, Mohr JP, Hier DB: The Stroke Data Bank: Design, methods, and baseline charac- teristics. Stroke 1988;19:547-554 6. Caplan LR, Hier DB, D'Cruz I: Cerebral embolism in the Michael Reese Stroke Registry. Stroke 1983;14:530-536 7. Mohr JP, Caplan LR, Melski JW, Goldstein RJ, Duncan GW, Kistler JP, Pessin MS, Bleich HL: The Harvard Cooperative Stroke Registry: A prospective registry. Neu- rology 1978;28:754-762 8. Schwartz B, Patil RS, Szolovits P: Artificial intelligence in medicine. Where do we stand? N Engl J Med 1987; 316:685-688 9. Spitzer K, Becker V, Thie A, Kunze K: The Hamburg Stroke Data Bank: Goals, design and preliminary results. / Neurol 1989;236:139-144 10. Richard RH: Evaluation of a medical data system. Comput Biomed Res 1970;3:415-425 11. Davis R, Buchanan B, Shortliffe E: Production rules as a representation for a knowledge based consultation program. Artiflntell 1977;7:15-45 12. Hudson DL, Estrin T: Derivation of rule-based knowledge from established medical outlines. Comput Biol Med 1984; 14:3-13 13. Clancey WJ: The epistemology of a rule-based expert sys- tem—A framework for explanation. Artif Intell 1983; 20:215-251 14. Shortliffe EH: Computer-Based Medical Consultations: MYCIN. New York, Elesevier, 1976 15. Szolovits P: Artificial Intelligence in Medicine. Boulder, Colo, Westview Press, 1982 16. Torasso P: Knowledge based expert systems for medical diagnosis. Stat Med 1985;4:317-325 17. Bouckaert A: Medical diagnosis: Are expert systems needed? IntJ Biomed Comput 1987;20:123-133 18. Szolovits P, Pauker SG: Categorical and probabilistic rea- soning in medical diagnosis. Artiflntell 1978;11:115-144 19. Chard T: Human versus machine: A comparison of a com- puter 'expert system' with human experts in the diagnosis of vaginal discharge. Int J Biomed Comput 1987;20:71-78 20. Zagoria RJ, Reggia JA: Transferability of medical decision support systems based on Bayesian classification. Med Decis Making 1983;3:503-509 21. 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 (Lond) 1986;11:75-81 22. Hier DB, Caplan LR, Hill H, Evens M, Sinha A: A microcomputer-based expert system to assist in localization of anatomic damage after stroke (abstract). Neurology 1984; 34(Suppl 1):83 23. Hayes-Roth F, Waterman DA, Lenat DB (eds): Building Expert Systems. Reading, Mass, Addison-Wesley, 1983 24. Mars NJ, Miller PL: Knowledge acquisition and verification tools for medical expert systems. Med Decis Making 1987; 7:6-11 25. Alvey PL, Myers CD, Greaves MF: High performance for expert systems: I. Escaping from the demonstrator class. Medrnf(Umd) 1987;12:85-95 26. Tuhrim S, Reggia JA: Feasibility of physician-developed expert systems. Med Decis Making 1986;6:23-26 27. Gross CR, Shinar D, Mohr JP, Hier DB, Caplan LR, Price TR, Wolf PA, Kase CS, Fishman IG, Calingo S: Interob- server agreement in the diagnosis of stroke type. Arch Neurol 1986;43:893-898 KEY WORDS • expert systems cerebrovascular disorders diagnosis D ow nloaded from http://ahajournals.org by on A pril 5, 2021