Expert Systems and Hypertext for teaching diagnostics

BETTINA I. REINHARDT

Department of Artificial Intelligence

University of Würzburg Am Hubland 97074 Würzburg, Germany

E-Mail: reinhard@informatik.uni-wuerzburg.de

Abstract: The use of computers in the field of higher education is common. Hypertext systems can present knowledge in a very powerful way to illustrate associations between information and to show visual information. A lot of intelligent tutoring systems are based on expert systems to retrieve the dynamic knowledge, necessary to criticize the student's actions. Some effort was done in combining these basic tools in different ways. In this paper we introduce a system, that integrates an expert system based tutoring system and a hypertext tool in a special way. The latter is used not only to present data, but also to criticize the student's recognition of the data. This is a very important part for classification problem solving, being based on the three skills: recognize the symptoms, interpret the symptoms and retrieve new symptoms. Another advantage of the system is the shell character, that allows experts to build new tutoring systems by themselves in a short time.

Introduction

Expert systems and hypermedia are successfully used tools in different domains (Fontaine, 1994; Silverman, 1992; and Gloor, 1991) and both can be considered as tools to present, influence and distribute knowledge. While hypermedia is mostly used to present data, expert systems mainly use data to resolve new data.

Expert systems are often used in diagnostics or classification, so we focus on this part of problem solving. Diagnostic skills can be divided in three parts:

* Recognizing symptoms:
for example, a medical student may know that a disease has some special skin problems, what does not mean, she or he can correctly recognize these skin changes during the examination of a patient. In flower classification the student may know that the flower has a specially formed calyx, but how does this really look like?

* Resolving diagnoses from the given symptoms:
that means the ability and knowledge to select a diagnosis, based on the given, correctly recognized symptoms, a diagnosis can be a disease or a flower, corresponding to the examples above.

* Asking for new symptoms to be become sure about the suspected diagnosis:
for example a new examination or a blood test of the patient; in the flower example it could be to cut the stem of the plant.

While there are some programs that show visualized symptoms and expect the student to answer questions about it, there are no tutoring systems with declarative knowledge representation integrating all three skills mentioned above. At a closer look some of these systems use hypermedia tools to improve their case presentation and therefore support the student's skills in recognizing symptoms, but none to our knowledge is able to criticize the student's recognition capabilities in a generic way. There are tutoring systems with a declarative knowledge representation, like Guidon (Clancey, 1990), helping the students to learn resolving diagnoses and how to ask for new examinations, but they do not handle the recognition of symptoms out of a visual presentation.

In the following, an introduction to the different ways to use hypermedia for computer based learning will be given. After introducing the underlying tools for our integrated system, the system itself will be presented. At the end there is a short conclusion and an outlook.

Hypermedia in tutoring systems

A hypermedia document can be used as a helpful part of a tutoring system or can be the tutoring system by itself. An example for a stand alone hypermedia teaching system is the lecture about hypermedia concepts (Gloor, 1991) or the business program HERMES (Pohl, 1995). While HERMES is used outside of the classroom, giving the students the opportunity to widen their lecture knowledge by exploring the HyperCard stack with pictures, text, sound and videos, the system by Gloor was used as the core of his lecture about hypermedia concepts at the University of Zurich. These systems present data to a student in a very attractive and organized way, but do never control the student's knowledge about the domain, besides the usual controls by a human teacher during the class.

Hypermedia can be used in tutoring systems in many variations. A good example is the Cardiac Tutor (Woolf, 1995) where hypermedia is used to visualize simulation. Nevertheless we want to focus on the abilities of hypermedia to improve teaching systems for classification problems with declarative knowledge bases.

There are different ways hypermedia can help teaching systems, based on a declarative knowledge base. Some are:

* presentation of facts, independent to the special case,

* case presentation in an independent hypermedia document,

* presentation of special symptoms in independent documents and

* case presentation integrated in the criticizing process.

One system, that works with the first method is NEUREX (Starita et al., 1995), a case oriented tutorial expert system for the diagnosis of neurogenic diseases of the lower limbs. It implements a process, based on anatomical, physiological, symptomatological knowledge and on the results of instrumental tests. While running, it explains its reasoning by interacting with the user in a friendly way and assists the user to reach possible diagnoses. NEUREX contains four hypermedia based atlases, where the students can explore nerve systems.

A tool for building such special multimedia-supported instructional systems is presented in (Kong, 1994). For example there are some applications teaching C++ Programming, LAN and Lisp.

The presentation of the case data in a hypermedia document is done in the system Abdominal Pain (Eitel, 1991), where one patient record is transferred to a sequence of pictures, texts and videos to explain the symptoms. The student can explore the case and the correct examination method. There are just a few patient cases with a lot of very entertaining hypermedia documents.

The presentation of single symptoms in an instructional system was chosen in (Hitzges, 1995), where an intelligent hypermedia system for problem solving in the domain of fault finding was built. The use of the system is to educate young mechanics to work with a complicated machine. Hypermedia documents, like pictures, texts and videos are used to illustrate single symptoms, like blinking control lights. There are also documents, that show how to repair the machine after finding the problem.

The system introduced in this paper is the only one that integrates the symptom recognition critic in a problem solving process. It is the combination of the two systems TRAINER (Reinhardt&Schewe, 1995) and HITS (Reinhardt&Remp, 1995). While the resolving of diagnoses and the retrieval of new data is performed and criticized by the tutoring system TRAINER, based on the expert system toolbox D3 (Puppe et al, 1994), the presentation of symptoms and the links from the information elements (pictures, text) to the knowledge base is done in the hypertext system HITS.

For this kind of hypermedia presentation it is not important to be hypermedia but it allows the student to bring in symptoms she or he is supposed to recognize, which can be criticized by the teaching system or checked upon with the real data behind it by the student him/herself.

The integration of the systems TRAINER and HITS is an example for a case presentation integrated in the criticizing process. While the hypertext system is used to illustrate the case or single symptoms, the tutoring system can criticize the student's recognition of these visual facts. In the moment there are two applications of this system running; one is a medical tutoring system about cardiological diseases, where the ECGs are shown to the student in pictures. An additional tool was implemented to help the students to measure the differences between the waves. The second application lies in the field of biology; students have to classify plants by given pictures of the plant, single parts of the plant and cuts. More applications in different domains are planed.

The basic tools

TRAINER

The tutoring system TRAINER is based on the expert system toolbox D3 (see table 1), that has a powerful graphical knowledge acquisition that makes it possible for domain expert to transform the knowledge in a knowledge base of D3 by themselves. It also has different kinds of problem solvers, where only the heuristic problem solver is used in the TRAINER. Another used component is the explanation tool, that visualize important parts of the knowledge base as well as dynamic problem solving knowledge of the actual case. TRAINER can use all knowledge and all case bases ever built in D3 to teach diagnostic skills.

There are two different ways to use the tutoring system; the first one, called "guided test" helps the student to understand the way diagnostic is done in this field by leading her or him through different steps of the problem solving process. For example in a medical domain the student has to study the history of a patient first, before she or he starts with some simple examination. They need to learn what examinations are important in the actual case. The guided test presents the patient data in exactly the right order, so that the student can get used to it. Every time he or she gets new patient data the student has to select suspected diagnoses and he or she has to justify the final diagnoses at the end of the case.

The second way to use the TRAINER is the "free test", where the student starts only with basic data and has to decide what examination has to be done next. Here the system can criticize not only the selected diagnoses of the students but also the set and order of the selected test. The student is not forced to justify a diagnosis or a chosen examination, but it is always possible to get an assessment of the selection and an assessment of the justifications.



Table 1: All major components of the expert system toolbox D3. The expert builds a knowledge base with the graphic acquisition tool. The user answers the questions of one of the dialogues, while one or more problem solvers choose diagnoses for the given problem. The explanation tool makes the process understandable for the user. Usually the expert uses the dialogue to build a case base, but every case ever solved by D3 can be stored. The Trainer uses only the heuristic problem solver (hypothize&test) as well as the knowledge and case base. It has its own dialogue and some parts of the explanation tool are used.

The main advantage of our system is the shell character, that makes it possible to build training systems with no additional costs out of given knowledge bases. The TRAINER is very well evaluated in several medical environments like rheumatology (Schewe et al., 1996), neurology (Puppe et al., 1995), cardiology (Reinhardt&Remp, 1995). There are more medical knowledge bases built in D3, that are not yet evaluated as training systems, but as consultation programs (Gappa, 1995; Puppe 1995).

HITS

The hypertext system HITS is a shell for building hypertext documents, highly corresponding to a knowledge base of the expert system tool box D3. It is divided in an author and a dialog mode. In the author mode the expert can connect hypertext elements like pictures, graphics or pure text and also diagnoses and symptoms of a knowledge base. Building such a hypertext document does not need any programming skills of the expert, because every action can be made in a graphical user interface.

The expert has to define a picture or a text as a start object. In pictures she or he can define areas as click sensitive similar to the handling in an ordinary drawing program. By changing the tool and clicking on the sensitive area a new pop up opens, where she or he can decide to connect an existing or a new picture, a text or an element of the knowledge base. If the expert chooses a knowledge base element a hierarchy of the element is opened. By existing pictures and texts, a list is opened for selection.

Table 2: Left there is a typical picture of a plant, where the expert can mark sections. A mouse click on the section will open a special menu, where the expert can link some elements to this section. Right is the template for the text window, that is transferred to a text window.

The click sensitive areas in text elements are defined in a special parentheses structure. The normal text is written down in a usual window, where a special symbol is around the chosen areas. By clicking a button in the window the text is transferred to a new window with the sensitive text, where a click on the text opens the same pop up as in the pictures.

A connection to a new element needs the name of the new window and a text, used in the dialog for open it, while a connection to an existing window only needs a new text. So the window can be named "calyx", while the text to open it is "zoom" out of one picture and "calyx" out of another picture. This helps to use the same window in different environments.

In the dialog mode the start window is opened and the connected windows can be opened with a pop up in the sensitive areas. In the settings the expert can decide if the areas in the pictures are framed or if only the cursor changes. Sometimes the frames are needed for beginners while the intermediate or advanced student had to find the areas by herself or himself.



Table 3: Presentation of a plant in HITS.

Integration

After introducing the two subsystems, how do they work together? The TRAINER is able to criticize the problem solving abilities of the student, while HITS can present a case in a realistic way. We decided to add abilities to criticize the student by classifying the symptoms of the given case. Because critic is the responsibility of the TRAINER, more knowledge and another form of assessment were added. When the student opens a new case, the TRAINER checks for a suitable HITS document. If there is one, the start window will open and the information that can be found in the document is filtered from the presented case data.

Now the student has two different ways of symptoms: the one the TRAINER presents him in the usual dynamic hierarchy and the one HITS presents him in the pictures and texts of the hypertext document. To criticize the classification of the visual symptoms there had to be a dialog to offer a way to fill them out in the ordinary case presentation (see table 4). Therefore the visual symptoms are integrated in the dynamic hierarchy and can be answered with a pop up. The answers are used for the problem solving process, so that the critic of the problem solving of the students is based on the same data.

The answers can be changed every time and in the free test a critic of the classified symptoms is available by clicking a button from the dynamic hierarchy, while in the guided test, the critic is available at every diagnoses' assessment. In the assessment of the symptoms the answers are divided in three categories: correct answered, incorrect answered and yet not answered.

In all cases the student can click on the symptom and look at the hypertext element where he or she should have seen the symptom and he or she can also get the correct answer (see table 5). At this point the problem solving critic is based on the incorrect answer, so the student has to change his input and think about the consequences to the selected diagnoses. So the student is able to follow the consequences of different symptoms in different combinations. It could be that some other symptoms generally point to the correct diagnoses, but an incorrectly classified symptom stopped the rule. So she or he can explore the inferences between the symptoms and get a deeper understanding of the problem solving process.



Table 4: The student sees the symptom in Hits and answers the unknown symptoms in the dynamic hierarchy of the Trainer.

Table 5: In the assessment window of the system, the recognized symptoms are divided in correct, incorrect and not recognized symptoms. Through a pop up from every symptom the student can see the correct answer or the hypertext element where she or he should see the symptom.

The integration of TRAINER and HITS is a system of the last group of instructional systems using hypertext. It shows the two way communication in a problem solving process. The tutoring system sends the message to open a document to a case, some symptoms or diagnoses to the hypertext system, while the hypertext system communicates with the instructional system, what symptoms should be recognized by the student and where he or she can see them.

Conclusion

Based on our knowledge, the integration of TRAINER and HITS is the only system realizing a control of the student's recognition capabilities in a problem solving process. Especially in the evaluation of the flower classification application the students, who worked with the system were highly motivated (Ernst, 1996). Usually they can only use flowers, that are from the region in the special time of the year. With the teaching system they could classify flowers from every time in the year in different growth phases from different regions better than in a conventional textbook.

While in the flower classification application it is absolutely necessary to have a hypertext document for each case constructed by the expert, in most of the possible applications a case presentation could be generated out of a given media base and its links to the database. For each knowledge base element one or more hypermedia elements could be stored and a template for cases could be designed, so that the system generates a hypermedia document for every case out of this template. The hypermedia elements could be selected with some additional restrictions, like gender of the patient, intensity of the disease. This would reduce the time to build such an integrated training system to a minimum.

In the moment HITS is just able to handle text and pictures, while sound and video are easy to add. Some work has to be done to acquire new data material from the expert's domain to make the case presentation more realistic. Another point is the integration of the information system, currently build to D3. So the textbook or some atlases can be used out of the instructional lesson.

References

Clancey W.J. (1990). Knowledge-Based Tutoring - The GUIDON Program, MIT Press Cambridge.

Eitel (1991). HyperCard based patient simulation with IVD. In Dtsch. Ärzteblatt 88, 2623-2626 (in German).

Ernst Roman (1996). Building a tutoring system for flower classification. Diploma thesis at the University of Würzburg, Department of biology.

Fontaine D., Le Beux P., Riou C., Jacquelinet C. (1994). An Intelligent Computer-Assisted Instructional System for Clinical Case Teaching. In Methods of Information in Medicine 1994,33, 433-445.

Gappa U. (1995). Graphical Knowledge Acquisition Tools. DISKI 100, infix-Verlag (in German).

Gloor P.A. (1991). Presenting Hypermedia Concepts using Hypermedia Techniques. In: Maurer H. Hypertext/Hypermedia '91.

Hitzges A. (1995). Problem oriented support by error diagnosis. In Schoop E., Witt R., Glowalla U.: Schriften zur Informationswissenschaft, Bd. 17, Universitaetsverlag Konstanz (in German).

Kong H.P. (1994). An intelligent multimedia supported instructional system. In: Expert Systems with Applications, Vol. 7, No. 3.

Poeck K., Tins M. (1993). An Intelligent Tutoring System for Classification Problem Solving. In Ohlbach: GWAI-92, LNAI 671, 210-220.

Pohl C. (1995). Navigation in a hypermedia teaching system. In Schoop E., Witt R., Glowalla U.: Schriften zur Informationswissenschaft, Bd. 17, Universitaetsverlag Konstanz (in German).

Puppe F., Poeck K., Gappa U., Bamberger S., Goos K. (1994). Reusable Components for a configurable Diagnostics Shell. In Künstliche Intelligenz 2/1994, 13-18 (in German).

Puppe F., Reinhardt B., Poeck K. (1995). Case Oriented Neurology Trainer. In Künstliche Intelligenz 95/1, 52-54. (in German).

Puppe B. (1995). Reflections on Building Medical Decision Support Systems and Corresponding Implementation in Diagnostics Shell D3. In Proceedings of the Fifth Conference on Artificial Intelligence in Medicine Europe , 1995, 282-292.

Reinhardt B., Remp T. (1995). An example for integration of hypertext and expert systems in cardiologic diseases. In Proceedings to GISI95, 275-283 (in German).

Reinhardt B., Schewe S. (1995). A shell for intelligent tutoring systems". In Proceedings to AIED 1995, 83-90.

Schewe S., Quack T., Reinhardt B., Puppe F. (1996). Evaluation of a Knowledge Bases Tutorial Program in Rheumatology - a Part of a Mandatory Course in Internal Medicine. In Proceedings to the Conference on Intelligent Tutoring Systems (ITS96) 1996.

Silverman B.G. (1992). Survey of Expert Critiquing Systems: Practical And Theoretical Frontiers. In Communications of the ACM April 1992/Vol.35,No.4.

Starita A., Majidi D., Giordano A., Battaglia M., Cioni R. (1995). NEUREX: a tutorial expert system for the diagnosis of neurogenic diseases of the lower limbs. In: Artificial Intelligence in Medicine 7 (1995) 25-36.

Warner H., Haug P., Bouhaddou O.; Lincoln M., Warner H. Jr., Sorenson D., Williamson J., Fan C. (1988). ILIAD As an Expert Consultant to Teach Differential Diagnosis. In SCAMC 1988 371-376.

Woolf B., Hall W. (1995). Multimedia Pedagogues:Interactive Systems for Teaching and Learning. In IEEE 1995, pages 74-80.

Acknowledgements

Thanks to all the experts who use TRAINER and HITS to build tutoring systems, who use these systems (even more important) and who help me with their comments improving the systems, especially Stefan Schewe, Klaus Poeck and Roman Ernst, who fought his way to all the phases of HITS without complaining too much. Also thanks to all my colleques and developpers of D3.