a prototype system for perinatal knowledge engineering

Sokol & Chik, Prototype system for perinatal knowledge engineering 273

J. Perinat. Med.16 (1988) 273

A prototype system for perinatal knowledge engineering using anartificial intelligence tool

Robert J. Sokol and Lawrence Chik

Department of Obstetrics and Gynecology, Wayne State University/Hutzel Hospi-tal, Detroit, Michigan, U.S.A.

1 Introduction

Expert systems are generally regarded as programsthat exhibit "intelligent," i.e., human-like attri-butes. These attributes typically would include theability 1) to conduct free-form dialogue, 2) toprovide explanations as to 'why' a response wasproduced when asked, 3) to accept new knowledgeand retract erroneous or obsolete facts and 4) toresolve problems which were NOT explicitly pre-programmed. Expert systems using artificial intel-ligence (AI) programming tools are becomingmore widely available for use in large corporatemainframe systems, as well as in microcomputingenvironments. Such systems are beginning to haveimpact in the business world. Some specializedapplications have been investigated in medicine aswell. The best known of the medical expert systemsinclude the MYCIN [1] and the INTERNIST [2]systems. These systems are speculated to haveinvolved 15 to 30 people years of developmenttime. Because such systems have not been portedto laboratory scale or, preferably, departmentalscale or personal computers, their impact on med-ical education and clinical medicine has been lim-ited.To the time of the First World Symposium onComputers in the Care of the Mother, Fetus andNewborn (March, 1987), there had been manymedical application programs termed "expert sys-tems" by their authors. Typically, these programsconsist of diagnostic and/or treatment algorithmsrepresented as explicit logical statements pro-grammed in standard languages, including Basicand Fortran. Thus, for example, the current au-thors reported the development and utility of sucha program for the diagnosis of normal and ab-

Curriculum vitae

ROBERT J. SOKOL, M. D. iscurrently Professor andChairman of the Depart-ment of Obstetrics and Gy-necology at Wayne StateUniversity and Chief of theDepartment in the DetroitMedical Center/HutzelHospital. With about 450publications, his major re-search interests includecomputer applications inperinatal medicine and the assessment of prenatal risks asdeterminants of infant outcome, particularly neurobeha-vioral development. Grant support is in the area of alcohol,as a perinatal risk.

normal labor progression 14 years ago [5]. Basedupon the patient's parity, the time of onset oflabor and the time, dilatation and station of eachvaginal examination, that program, still in use,generates any of 53 appropriate diagnostic mes-sages, singly or in combination. This type of ap-plication program can serve very useful purposesat the bedside for clinical care, educationally andas a research tool. Though other "expert systems"have been reported by perinatologists, to the bestof our knowledge, the "intelligent" attributes ofmodem AI programming tools have not been usedin significant applications in perinatal medicine.Yet, the potential for "an intelligence machine"remains conceptually promising, and ever more sowith progressively more powerful departmentalcomputers and workstations.

1988 by Walter de Gruyter & Co. Berlin · New York

274 Sokol & Chik, Prototype system for perinatal knowledge engineering

Our goal in the current study was to develop auseful perinatal knowledge base and inference en-gine for perinatal consultation at the level of thesenior resident and perinatal fellow. This manu-script describes our initial experience in attemptingto apply an artificial intelligence programmingtool, 'PROLOG,' with a focus on the problemsencountered in using the standard "bottom up"approach to building a knowledge base. Finally,we will describe a "top down" approach, i.e.,global to the development of a perinatal knowl-edge base and inference engine and the resultingprototype system using a variety of system utilitiesin conjunction with PROLOG.

2 A "bottom up" approach

Inasmuch as the current investigators had no ex-perience in developing artificial intelligence-basedexpert systems, we decided to employ the textbookapproach and mimic the process described in thedevelopment of such systems in industry. Typi-cally, there are two participants necessary for thedevelopment of such systems. The first is the "ex-pert," i.e., the individual whose brain is to be"picked" for facts and rules for the knowledgebase to be developed. In the current situation, thisindividual was termed the perinatal expert. Therole of this individual was to identify a relevantarea to specify specific topics and problems andto describe perinatal risks, appropriate medicalinterventions, including both workup and treat-ment, and to define expected perinatal outcomes,as well as to provide explanations for the variousrelationships included in the knowledge base.

In developing an expert system, the second nec-essary individual is the knowledge engineer. Thisindividual's role includes first, developing an ap-propriate framework for facts and decision logic.Second, he/she must design a logic program fordecision support and develop a program environ-ment for both the interface with the user and theexpert, i. e., for the consultative use of the system.Finally, the knowledge engineer must be con-cerned with the design of a natural language in-terface for both input and output from the system.

For development of our prototype system, a VAX11/750 was available. The operating system is theUniversity of California at Berkeley 4.1 BSD Unix[6]. After an evaluation of available AI program-

ming languages, we chose the University of NewHampshire NH Prolog. Following primarily thePOPRAS [4] uniform perinatal record, we choseto experiment with a limited risk intervention out-come model, using the framework — risk, workup,treatment, outcome and explanation. An exampleof the framework and a single entry within thisframework is shown in figure 1.

- UNH PROLOG 1.3 ··{initialization}

WHAT IS THE TOPIC THAT YOU ARE INSTERESTED IN ? I need information on iugr.

PLEASE WAIT, I AM LOOKING FOR SIMILAR TERMS.

SEARCH LIST:*[IUGR]*[FETAL,GROWTH,RETARDATION]

** SEARCHING FRAMES **

PROBLEM TYPE: POOR OB HISTORY

*[INTRAUTERINE,GROWTH,RETARDATION]«[GROWTH,RETARDATION]

SYMPTOM:WORKUP:

PREVIOUS SGAEARLY DETECTION BY SERIAL MCDONALD MEASUREMENTS ORULTRASOUNDS, NSTS IN THIRD TRIMESTERIUGR FETUSRISK TENDS TO RECUR BASED IN PARTON PERSISTENT CAUSES SUCH AS CHRONIC HYPERTENSION,SMOKING OR SUBSTANCE ABUSE

** SEARCHING TEXT **

INTRAUTERINE GROWTH RETARDATION AND CONGENITAL ANOMALIES ALSO CONTRIBUTE TONEONATAL MORTALITY. (5)

SPECIFIC ANTEPARTUM RISKS FOUND TO BE ASSOCIATED WITH NEONATAL MORBIDITY INCLUDEMATERNAL CARDIOVASCULAR - RENAL COMPLICATIONS , IN PARTICULAR , PREGNANCY ·INDUCED HYPERTENSION , DIABETES MELLITUS, CLINICALLY SUSPECTED INTRAUTERINEGROWTH RETARDATION. AND OTHER SEVERE MEDICAL / OBSTETRIC COMPLICATIONS.

FURTHERMORE , THE COMBINATION OF CHRONIC AND PREGNANCY - INDUCED HYPERTENSION ,IS ASSOCIATED WITH A CONSIDERABLE INCREASE IN RISK FOR INTRAUTERINE GROWTHRETARDATION AND ABRUPTIO PLACENTA.

CYANOTIC HEART DISEASE IS A CLEAR RISK FOR INTRAUTERINE GROWTH RETARDATION. (25)

ALCOHOL IS A WELL ESTABLISHED TERATOGEN AND IS ASSOCIATED WITH INTRAUTERINEGROWTH RETARDATION. AS WELL AS FETAL ALCOHOL SYNDROME AND ASSOCIATED MENTALRETARDATION. (30)

ACUTE FATTY LIVER OF PREGNANCY CONTINUES TO HAVE A HIGH MORTALITY RATE FOR THEMOTHER. (37) STOMACH STAPLING, A TREATMENT FOR OBESITY WHICH HAS FALLEN OUT OFFAVOR MAY SOMETIMES STILL BE ENCOUNTERED AND HAS BEEN ASSOCIATED DURINGPREGNANCY WITH AN INCREASED RISK OF INTRAUTERINE GROWTH RETARDATION AND WITH THERISK FOR MATERNAL HEPATIC NECROSIS . (38)

PERHAPS , THE MOST COMMONLY MISSED PROBLEM ASSOCIATED WITH THE UTERINE SIZEBEING SMALLER THAN EXPECTED , IS INTRAUTERINE GROWTH RETARDATION. THE CLINICALINTERPRETATION BEING THAT THE PATIENT HAS NOT PROGRESSED AS FAR IN PREGNANCY ASSHE HAD THOUGHT .

DO YOU WANT TO CLEAR THE INPUT FOR RECURSION <Y/N>? yDO YOU WANT A PHRASE LIST FROM THE OUTPUT FOR RECURSION <Y/N>? y

Figure 1. Edited printout for prototype perinatal consul-tation session. The user input follows the T and theblank lines between paragraphs in 'searching text' implyuser options to continue. Numbers in parentheses indi-cate reference in on-line bibliography.

J. Perinat. Med. 16 (1988)


After, perhaps, 50 person hours of work, it "rap-idly" became evident that this task was not goingto be as easy as it first looked. In normal pro-gramming, a set of well-defined input/output spec-ifications is passed to a programmer, who thenimplements those specifications in a set of programstatements. In building a knowledge base, the "ex-pert" must participate throughout the process.Furthermore, it became clear that the breadth ofthe task we had chosen, i. e., to develop a perinatalconsultant was much too broad. Either, we wouldneed to change our overall goal or our approach.We chose the latter route.

3 A "top down" approach

A straightforward description of the top downapproach we envisioned was one which is morelike doing a sculpture than building a house. Itwas readily apparent to us that there is a vastamount of information available, first, from theliterature and second, from service statistics fromperinatal data bases. We, therefore, focused ourefforts on developing methods for the appropriateorderly retrieval of such information, using AItools where relevant. We determined that for theprototype under development, there would be twoprimary sources of knowledge. The first of thesewas the service statistics for one year's experienceon the Obstetric Service at Hutzel Hospital/WayneState University. These statistics include the fre-quencies of antepartum and intrapartum risks andcomplications, as well as variables covering im-mediate neonatal outcome. The second source of"perinatal knowledge" was a manuscript we hadrecently prepared for a resident level textbook [3].This manuscript covered, in broad general outline,the diagnostic methods and treatment of a fullrange of high risk pregnancy complications. Themanuscript was 30 double-spaced typed pages andincluded 715 sentences in 106 paragraphs, exclud-ing references.Accessing information from the service statisticswas straightforward and will not be described herein detail. Rather, we focus on the complex issueof extracting facts from a manuscript to build aperinatal knowledge base. Using the chapter as abasis and after trying several approaches, we cameto the conclusion that the UNIX writing styleanalysis utility 'style' could be useful in retrievingsome information from the manuscript by lookingfor the parts of speech.

4 Knowledge extraction from a manuscript:methods

For this study, two methods of extracting knowl-edge from the literature, here represented by asingle chapter, broadly covering perinatal issues,were compared. We used the first method as our"gold standard" and termed this method "semi-handmade." The manuscript was processed intoone sentence per line using UNIX utilities 'sed'and 'nroff and parsed for the parts of speechsentence-by-sentence using 'style.' A Fortran pro-gram was written to compile the parts of speechto produce a list of key words and phrases (nounsand noun phrases). The resulting list was displayedon a CRT and edited by the perinatal expert, withthe help of the knowledge engineer. This was rel-atively simple, inasmuch as the task was limitedto eliminating words and phrases which, for thepurpose at hand, did not identify important peri-natal concepts.The second method was termed "semiautomated."In this method, the manuscript was "spelled" firstusing a medical dictionary to validate the correctspelling, and then using a dictionary without med-ical words. As the output of the spelling programtypically includes misspelled words and words un-known in the dictionary, the use of a nonmedicaldictionary allowed us to simply identify medicalterms and the corresponding sentences in the con-text.The results of the "semihandmade" and "semi-automated" methods of identifying key terms areshown in table I. It can be seen that using the"semihandmade" approach, after editing, therewere 664 key nouns and phrases representing per-inatal concepts, i. e., risks, interventions and out-comes. In comparison, using the spell program,approximately half as many terms were identified.Our next step was to examine and compare thesuccess of our two methods of extracting infor-mation from the manuscript. The criterion func-tion chosen a priori was the extraction of pertinentsentences representing medical knowledge fromthe manuscript. The results of this comparison areshown in the contingency table in table II. Usingthe gold standard "semihandmade" method, keysentences were taken as including greater than orequal to two key nouns and/or phrases. Of the715 sentences in the manuscript, 30% (213) wereidentified as key by the experts. Of these 213, 169were also identified by the semiautomated methodusing the spell check program (sensitivity = 79%).



Table I. Analysis of nouns and noun phrases by 'style'analysis (semihandmade) and analysis of medical termsby 'spell' (semiautomated) in manuscript

Number Uniq-sort Edited

SemihandmadeKey nounsNoun phrasesTotal

SemiautomatedMedical terms

16731780

8151332

312

241423664

Based on these results, we concluded that the spellprogram represented a useful step of selecting keysentences from this chapter manuscript. Consider-ing that this manuscript was reasonably repre-sentative of other materials in the perinatal area,e.g., using very similar terms, we speculate thatthis approach may be useful for "noise disposal"in perinatal knowledge acquisition. However, thecomplexity of natural language will clearly neces-sitate better understanding of grammar rules andsemantics for more precise information retrievalor "knowledge extraction".

Table Π. Correlation of 715 key sentences by semi-handmade method and semiautomated methods

SemihandmadeKey Othersentences sentences

Semiautomated1 or more medicalterms in sentences

169 146 315

no medical termsin sentences

44

213

356

502

400

715

Sensitivity: 79%Specificity: 71%Positive productive value: 54%Negative productive value: 89%Chi-square: 153.3p < .001Adjusted R2: 39.8%

Twenty-one percent of the key sentences were notidentified. Of 502 sentences not judged to containkey information by the experts, the semiautomatedmethod, based on the lack of a medical word inthe sentence, correctly identified 356 (specificity= 71%).The semiautomated method chose 315 sentencesof the 715 as important, based on the presence ofat least one medical term. Of these 315, 169 werejudged to be key by the experts for a positivepredictive value of 54%. Similarly, of the 400sentences that did not have a medical word, 356were judged by the experts not to be key, for anegative predictive value of 89%. Calculated fromchi-square and the marginals of the contingencytable, as a measure of association, the adjusted R2

was 39.8%.

5 A prototype system for knowledge engineeringHaving developed a potentially practicablemethod for knowledge extraction, we developedan overall concept of a prototype system forknowledge engineering. This is shown graphicallyin figure 2. Using text processing utilities fromUNIX, information may be extracted from one ormore manuscripts in a given perinatal area. Keysentences are then collected and coded intoPROLOG compatible files as facts and rules,yielding a knowledge base. An inference engine,which could invoke system utilities, allows theexpert and the knowledge engineer to interact withthe system to 'assert' new facts and 'retract' er-roneous and obsolete facts, as well as allowing theuser to consult the knowledge base. In the pro-totype perinatal expert system being reported here,the knowledge base includes four componants:

1) A similar term list. We have found this nec-essary to make it possible for the computerto identify all appropriate areas within theremainder of the knowledge base. Thus, forexample, the computer must "know" thatsmall-for-gestational age and intrauterinegrowth retardation, as well as fetal growthretardation, represents similar concepts.

2) Service statistics, as described above.3) Key sentences as extracted by the methodol-

ogy described above. These are expressed asPROLOG "facts" with appropriate syntax.

4) Finally, the full manuscript, along with itsreferences, is included. This provides a simplesolution to the "natural language interfaceproblem" on the output side of the loop. Themachine responds in natural language pre-pared by an expert.



K N O W L E D G EB R S E

Figure 2. A prototype system for knowledge engineering.

only the intercepts, i. e., areas in the knowledgebase in which IUGR is related to hypertension.Under interactive control, it is possible to branchfrom, say the first identification of IUGR in theknowledge base, to other key terms found in sen-tences mentioning IUGR, hypertension, or both.This latter approach allows for the rapid assess-ment of knowledge within a broad perinatal con-text.At the end of run, the retrieved material can beanalyzed to produce a new list of relevant topicsfor recursive search. The entire process emulatesthe response of an instructor to an inquisitivestudent.

The inference engine in our prototype system in-cludes preliminary rules and meta rules for anintelligent history taking program modeled afterthe POPRAS forms. In addition, portions of arecursive literature search program, usingPROLOG, have been implemented.An edited printout for a consultation with thisprototype is shown in figure 1. To consult thesystem, the user might proceed as if he were usinga medical book. Instead of using the table ofcontents or index, he/she might enter a key wordor noun phrase of interest, e. g. iugr. Alternately,he/she might conduct a dialogue with the terminaland enter "I need information on intrauterinegrowth retardation." Using the parsing routine,the program identifies the key term, i.e., intra-uterine growth retardation, and then proceeds se-rially through the frame structured knowledgebase (risk, intervention and outcome) and the serv-ice statistics which include information on thefrequency of diabetes mellitus and growth retar-dation and their relationship to perinatal outcome.Next, the key sentences are evaluated and appro-priate segments of the entire manuscript, alongwith references, may be output under interactivecontrol.We are experimenting with various types of outputand the order of output to determine what willmost likely be of most use to the individual con-sulting the knowledge base. It may be noted thatsometimes an in-depth search concerning a singlesubject is most appropriate, with an exhaustivesearch for everything in the knowledge base, e. g.,concerning intrauterine growth retardation, priorto looking for hypertension or other related con-cepts. Alternativety, the user may wish to identify

6 Further steps

Based on our results to date, it is clear that con-siderably more work on the natural language in-terface for query input is necessary for free formdialogue. The relatively simple grammar rules of'style' limit acceptable dialogue to relatively simplesentences only.So far, we have only evaluated our semiautomatic"fact" extraction methodology on a single manu-script. Other manuscripts will need to be proc-essed, using this methodology to obtain a betterestimate of its utility. Also, considerably morework on the engine, including both the input/output interface, is needed to determine the mostuseful strategies for effective consultation, i.e.,including interactive choice for depth versusbreadth search paths.

7 Speculations and conclusions

Based on the experience gained over the past yearand the results presented in this paper, it appearsthat a generalized approach to reducing the sizeof the immense task of perinatal knowledge en-gineering may be attainable. If so, it will undoubt-edly be more efficient than hand building a mul-titude of individual knowledge bases of very lim-ited breadth. Since the prototype system reportedhere functions in a computing environment withthe capacity of a contemporary workstation, per-haps, practical educational and clinical applica-tions might be possible. It is clear that an AI tool,such as PROLOG has the versatility to deal withmany aspects of the list processing applications;



but it is not a necessary tool, in that other systemutilities and conventional programming tools canbe employed to accomplish the some goal. Basedon the experience reported here, it is reasonable

to suggest that most elaborate expert system de-signs will require a combination on AI and con-ventional programming tools for design and de-velopment.

Abstract

Though several perinatal expert systems are extant, theuse of artificial intelligence has, as yet, had minimalimpact in medical computing. In this evaluation of thepotential of AI techniques in the development of acomputer based "Perinatal Consultant," a "top down"approach to the development of a perinatal knowledgebase was taken, using as a source for such a knowledgebase a 30-page manuscript of a chapter concerning highrisk pregnancy. The UNIX utility "style" was used toparse sentences and obtain key words and phrases, bothas part of a natural language interface and to identify

key perinatal concepts. Compared with the "gold stand-ard" of sentences containing key facts as chosen by theexperts, a semiautomated method using a nonmedicalspeller to identify key words and phrases in contextfunctioned with a sensitivity of 79%, i. e., approximately8 in 10 key sentences were detected as the basis forPROLOG, rules and facts for the knowledge base. Theseencouraging results suggest that functional perinatal ex-pert systems may well be expedited by using program-ming utilities in conjunction with AI tools and publishedliterature.

Keywords: Artificial intelligence, clinical decision support, data base, expert systems, perinatology, PROLOG,UNIX.

Zusammenfassung

Ein Systemprototyp in Form eines Werkzeugs der künst-lichen Intelligenz für perinatales Ingenieurwesen in derWissenschaftExpertensysteme werden allgemein als Programme ver-standen, die „intelligente", d.h. menschenähnliche At-tribute aufweisen. Diese Attribute beinhalten bezeich-nenderweise die Fähigkeit,1) freie Dialoge zu führen2) bei Anfrage Erklärungen zu liefern, „warum" eineReaktion erfolgte,3) neue Erkenntnisse zu übernehmen und fehlerhaftebzw. überholte Tatsachen zu verwerden und4) Probleme zu lösen, die nicht explizit vorprogrammiertwurden.Expertensysteme, die mit Werkzeugen der künstlichenIntelligenz arbeiten, verbreiten sich zunehmend in derGeschäftswelt, hatten jedoch, obwohl einige spezielleAnwendungen in der Medizin erforscht worden sind,bisher geringfügigen Einfluß auf die medizinische Aus-wertung.Es wurden bisher eine Reihe von medizinischen Anwen-dungsprogrammen unter der Bezeichnung „Experten-systeme" vorgestellt, jedoch nach den Kenntnissen derAutoren sind „intelligente" Attribute moderner AI Pro-grammierwerkzeuge bisher noch nicht bei wesentlichenAnwendungen in der perinatalen Medizin zur Anwen-dung gekommen. Das Ziel der vorliegenden Arbeit wares deshalb, eine brauchbare perinatologische Wissens-basis und eine Entscheidungshilfe für perinatale Bera-tung auf der Ebene eines Oberarztes und eines Assistenz-arztes zu entwickeln.Anfangs versuchten nur ein „Experte auf dem Gebietder Perinatologie" und ein „Wissensingenieur", ein klas-

sisches AI-System dadurch zu entwickeln, daß sie einGerüst einer Risikoaufbereitung, Behandlungsergebnisseund Erläuterungen aufstellten. Nach etwa 50 Arbeits-stunden an diesem Prototypsystem wurde deutlich, daßdiese Aufgabe nicht so leicht zu lösen war, wie es zu-nächst erschien. Ein besonderes Problem bestand darin,daß während des Aufbaus der Wissensbasis der „Ex-perte" die ganze Zeit hindurch am Prozeß teilnehmenmuß.Zusätzlich wurde klar, daß die gestellte Aufgabe, näm-lich Entwicklung eines Beraters für perinatologische Fra-gen, viel zu weit gefaßt war. Deshalb wurde ein andererWeg beschritten.Da umfangreiches Informationsmaterial sowohl aus derperinatalen Literatur als auch aus perinatologischen Da-tenbanken von Seiten der Dienstleistungsstatistiken zurVerfügung steht, konzentrierten sich die Bemühungenauf die Entwicklung von Methoden zur zweckmäßigenund folgerichtigen Auffindung dieser Informationen,ggf. unter Verwendung der AI-Werkzeuge. Als Grund-lage für das Prototypsystem wurden die Dienstleistungs-statistiken einer großen geburtshilflichen Klinik über einJahr und ein kürzlich erstellter Entwurf eines grundle-genden Lehrbuchkonzeptes über Hochrisikoschwanger-schaften verwendet. Dieses Lehrbuchkonzept bestandaus etwa 30 zweizeilig gedruckten Seiten mit 715 Sätzen(ohne Literaturhinweise) in 106 Paragraphen.Bezüglich der Problemstellung, „Wissen" aus der Lite-ratur herauszuziehen, d.h. aus einem bestimmten Ka-pitel, wurde ein „halbautomatisches" System mit demWissen, das die Experten aus diesen entsprechendenKapiteln herausgezogen haben, verglichen. Das a priorifestgelegte Auswahlkriterium war das Auffinden der



Sätze im Lehrbuchkonzept, die das medizinische Wissenrepräsentierten. Das halbautomatische System arbeiteteauf einer rechnergestützten elektronischen „Buchstaben-basis", indem es ein Wörterbuch ohne medizinischeFachwörter verwandte, was damit gleichzusetzen ist, daßdie medizinischen Ausdrücke einfach in den entspre-chenden Sätzen im Gesamttext ausgesucht wurden. Vonden 750 Sätzen des Lehrbuchkonzeptes wurden 30%von den Fachkräften als Schlüsselsätze identifiziert. Vondiesen 213 Sätzen wurden 169 (Sensitivität = 79%) vondem halbautomatischen System durch das Buchetaben-vergleichs-Programm identifiziert. Entsprechend wurden21% der Schlüsselsätze nicht gefunden. Von den 502verbliebenen Sätzen, denen von den Fachkräften keineSchlüsselinformation zugeordnet worden war, hat dashalbautomatische Verfahren, allein gestützt auf nicht-medizinische Wörter, 356 (Spezifität = 71%) richtig er-kannt.Auf der anderen Seite hat die halbautomatische Me-thode, ausgehend von wenigstens l medizinischen Aus-druck, von den 715 Sätzen 315 als richtig ausgewiesen.Von diesen 315 Sätzen wurden 169 von den Fachkräfteneine Schlüsselrolle, was einem positiven Vorhersagewertvon 54% entspricht, zugeordnet, Entsprechend wurde356 Sätzen von den verbliebenen 400 Sätzen, die keinmedizinisches Wort enthielten, von den Fachkräftenkeine Schlüsselrolle zuerkannt, was einem negativen Vor-hersagewert von 89% entspricht. Dieses Ergebnis führtzu dem Schluß, daß das Buchstabenvergleichsprogrammeinen hilfreichen Schritt bei der Auswahl von Schlüssel-sätzen aus diesem in Kapitel unterteilten Lehrbuchkon-zept darstellt.In dem hier vorgestellten perinatologischen Prototyp-system wurden die nachfolgend aufgeführten 4 Kom-ponenten als Wissensgrundlage eingeschlossen:1) eine Liste ähnlicher Ausdrücke, über die es dem Com-puter ermöglicht wird, zugehörige Bereiche im verblei-benden Rest der Wissensgrundlage zu suchen,2) Dienstleistungsstatistiken

3) Schlüsselsätze, die durch die oben beschriebene Me-thode herausgezogen, die in entsprechender Syntax alsPROLOG-Tatsachen dargstellt wurden, und4) schließlich das komplette Lehrbuchmanuskript mitseinen Referenzen, das bezüglich der Ausgabenseite derKette eine einfache Lösung des Problems der Verbindungzur menschlichen Ausdrucksweise darstellt.Die Maschine antwortet in einer vom Experten erstelltenmenschlichen Sprache. Die maschinelle Entscheidung in-nerhalb dieses Prototypsystems schließt vorläufige undhochbewertete Vorschriften für eine intelligente Über-nahme der Krankengeschichte, aufbereitet gemäß PO-PRAS-Formblättern, ein. Darüber hinaus wurden Teileeiner rekursiven Literatursuche unter dem ProgrammPROLOG implementiert. Wird dem System eine Bera-tungsanfrage gestellt, so wird diese vom System so in-terpretiert, als würde der Lehrer dem wißbegierigenSchüler antworten.Es wurde dargelegt, daß das beschriebene System einPrototyp ist, das noch wesentlicher Arbeiten bezüglichder Verbindung zur menschlichen Ausdrucksweise beiAnfragen und ebenfalls noch einer Entwicklung bezüg-lich der halbautomatischen Extraktion von Fakten ausder Literatur bedarf. Nichtsdestoweniger erscheint es,wie die hier vorgelegten Erfahrungen zeigen, vernünftigzu sein, daß ein derartiger Ansatz nicht nur möglich,sondern auch praktisch, erzieherisch und klinisch ver-wertbar ist. Ein AI-Werkzeug, wie PROLOG es darstellt,hat in vielfacher Hinsicht die Flexibilität zur Listenver-arbeitung; aber es ist auf der anderen Seite kein notwen-diges Werkzeug, das in andere Systembibiliotheken undkonventionelle Programme eingebaut werden müßte, umdieselben Ziele zu erfüllen.Ausgehend von den hier berichteten Erfahrungen er-scheint die Annahme als vernünftig, daß die weitausgrößte Zahl von Entwürfen für Expertensysteme AI undkonventionelle Programmierwerkzeuge sowohl beimAufbau als auch bei der Weiterentwicklung Bedeutunghaben sollte.

Schlüsselwörter: Datenbank, Expertensysteme, klinische Entscheidungshilfe, künstliche Intelligenz, Perinatologie,PROLOG, UNIX.

Resume

Prototype pour l'organisation des connaissances perina-tales untilisant un instrument d'intelligence atrificielleOn considere generalement les systemes experts commedes programmes qui paraissent «intelligente», c'est-ä-dire qu'ils possedent des attribute humains. Ces attributecomprennent typiquement la capacite 1) de conduire undialogue de forme libre, 2) de fournir des explicationstelles que «pourquoi» une reponse a ete effectuee lors-qu'elle a ete demandee, 3) d'accepter de nouvellesconnaissances et de desavouer des faits errones ou ob-soletes, et 4) de resoudre ce qui n'a pas ete explicitementpre-programme. Des systemes experts utilisant des outilsde programmation d'intelligence artificielle (IA)deviennent plus largement disponibles dans le monde du

travail, mais, bien que certaines applications specialiseesaient ete explorees dans le champ de la medecine, ilsnont pour Pinstant qu'un impact minime en informa-tique medicale.De nombreux programmes d'application medicale qua-lifies de «systemes experts» par leurs promoteurs ontete proposes, mais a la connaissance des auteurs, lesattribute «intelligente » des outils de programmation de

moderne n'ont pas encore ete utilises dans desapplications eignificativee en medecine perinatale. Le butde cette etude a ete de developper une baee de connais-sances perinatales utiles et une machine deductive pourla coneultation perinatale du niveau de l'interne experi-mente (eenior reeident) et du perinatal fellow.



Dans un premier temps, «un expert en perinatologie»et « un Ingenieur de la connaissance » se sont efforces dedevelopper un Systeme «classique» d'IA, ä l'aide d'uncadre d'elaboration de risque, de traitement de 1'evolu-tion et d'explication. Apres environ 50 heures de travailsur ce Systeme prototype, il est devenu evident que latäche ne serait pas aussi aisee que ce qu'elle apparaissaitde prime abord. Un probleme particulier a ete le suivant:1'expert doit participer tout au long du processus d'ela-boration de la base de connaissance. En outre il estapparu clairement que Fampleur de la täche choisie,c'est-a-dire mettre au point un consultant en perinato-logie, etait bien trop vaste. C'est pourquoi on a adopteune approche alternative.Puisqu'il existe une grande quantite d'informations dis-ponibles, en provenance de la litterature perinatale etdes statistiques des services tirees des bases de donneesperinatales, on a centre les efforts sur la realisation demethodes de restitution appropriee et classe de tellesinformations en utilisant les outils d'IA le ou ils sontappropries. Pour le Systeme prototype, on a utilise lesstatistiques d'une annee d'activite d'un grand serviced'obstetrique et un manuscrit recent redige pour unmanuel sur les grossesses a haut risque s'adressant auxresidents. Le manuscrit etait de 30 pages dactylogra-phiees a double interligne et comprenait 715 phrasesreparties en 106 paragraphes, ä Fexception des refe-rences.En se focalisant sur le probleme de l'extraction «desconnaissances» de la litterature, representee par ununique chapitre, on a compare un Systeme « semi-auto-mate » aux connaissances extraites de ce chapitre par lesexperts. Le critere choisi a priori etait l'extraction dephrases pertinentes representant les connaissances me-dicales du manuscrit. Le Systeme semi-automate fonc-tionne au moyen d'un « alphabet» informatise, utilisantun dictionnaire sans termes medicaux qui procure unemethode d'identification simple des termes medicaux ausein des phrases correspondantes au contexte. Parmi les715 phrases du manuscrit, 30% ont ete identifiees commedes cirteres par les experts. Parmi ces 213 phrases, 169ont ete identifiees par la methode semi-automatique uti-lisant le programme de verification alphabetique (sen-sibilite = 79%). Vingt et un pour cent des phrases cir-teres n'ont pas ete identifiees. Sur les 502 phrases quin'ont pas ete jugees contenir des informations criterespar les experts, la methode semi-automatique, fondeesur l'absence de termes medicaux dans la phrase, en aidentifle correctement 356 (specificite = 71%). La me-thode semi-automatique recommit 315 phrases parmi les715 comme importantes, en se fondant sur la presence

d'au moins un terme medical. Parmi ces 315 phrases, lesexperts en ont juge 169 comme etant des criteres, avecune valeur predictive positive de 54%. De la meme facon,parmi les 400 phrases qui ne contiennent pas de termesmedicaux, les experts ont juge que 356 n'etaient pas descriteres, avec une valeur predictive negative de 89%. Ense fondant sur ces resultats, on a conclu que le pro-gramme alphabetique representait une etape utile pourselectionner les phrases cirteres de ce manuscrit.Dans le Systeme perinatal prototype rapporte ici, la basede connaissance comprend quatre composantes:1) une liste des termes similaires, necessaire pour que1'ordinateur puisse identifier toutes les zones approprieesau sein du reste de la base de connaissances; 2) desstastiques du service; 3) des phrases criteres, extraitespar la metholdologie decrite ci-dessous, exprimees entant que «faits » PROLOG avec une syntaxe adequate,et 4) enfin, la totalite du manuscrit, y compris ses refe-rences, ce qui assure une solution simple au « problemede 1'interface avec le langage naturel» au niveau de lasortie de la boucle. La machine repond en langage na-turel qui a ete prepare par un expert. L'instrument de-ductif de ce Systeme prototype comprend des regiespreliminaires et des regies pour une histoire intelligenteprenant un programme presente apres les formes PO-PRAS. En outre, des fragments de programme derecherche de litterature recursive, utilisant le PROLOGont ete ajoutes. Untilise comme consultation, le proces-sus dans son ensemble utlisant ce prototype imite lareponse d'un instructeur ä un etudiant inquisiteur.On insiste sur le fait qu'il s'agit d'une Systeme prototypesur lequel un travail plus important est necessaire tantau niveau de 1'interface avec le langage naturel au niveaude l'entree de questionnement, qu'au niveau de l'extra-ction au sein de la litterature de «faits» devaluationsemi-automatique. Neanmoins, en se fondant sur 1'ex-perience rapportee ici, il est raisonable de suggerer quePutilisation d'une teile approche rend bien possible desapplications pratiques, educationelles et cliniques. Unoutil d'IA, tel que le PROLOG, a une grande souplessed'emploi pour coincider avec les nombreux aspects desapplications des processus de listing; mais ce n'est pason outil necessaire en ce sens que d'autres systemes etd'autres outils de programmation conventionellepeuvent etre utilises pour accomplir le meme objectif.En se fondant sur Pexperience rapportee ici, il est rai-sonable de suggerer que de nombreux types de systemesexperts necessiteront l'association de et d'outils deprogrammation conventionelle pour leur conception etleur developpement.

Mots-cles: Aide ä la decision clinique, base de donnees, intelligence artificielle, perinatologie, PROLOG, Systemeexpert, UNIX.



References

[1] BARR A, EA FEIGENBAUM: The Handbook of Ar-tificial Intelligence, Vol. II. HeurisTech Press, Stan-ford, CA, (1982) 184

[2] BARR A, EA FEIGENBAUM: The Handbook of Ar-tificial Intelligence, Vol. II. HeurisTech Press, Stan-ford, CA, (1982) 197

[3] BHATIA R, RJ SOKOL, M PERNOLL: Detection ofHigh Risk Pregnancy. In: PERNOLL M, BENSON RC:Current Obstetric and Gynecologic Diagnosis andTreatment, 6th Edition. Lange, Los Altos, CA,(1987) 246-254

[4] HOBEL CJ: Development of POPRAS (Problem Ori-ented Perinatal Risk Assessment System). In: BAHRJP (ed): Proceedings First National Conference onPerinatal Data Systems, 1980. Wisconsin Assoc forPerinatal Care, Medical College of Wisconsin

[5] SOKOL RJ, RS NUSSBAUM, L CHIK, M ROSEN: Com-puter Diagnosis of Labor Progression: I. Develop-ment of an On-Line Interactive Digital ComputerProgram for the Diagnosis of Normal and Abnor-mal Cervical Dilatation Patterns. J Reprod Med 11(1973) 149

[6] Unix Programmer's Manual, 4.1 BSD User Docu-ment, Computer Science Division, Department ofElectrical Engineering and Computer Science, Univ.of California, Berkeley, CA, 1981

Robert J. Sokol, M. D.Department of Ob/GynHutzel Hospital4707 St. AntoineDetroit, MI 48201, U.S.A.

a prototype system for perinatal knowledge engineering

Documents