Frank Puppe1, Bettina Reinhardt1, Klaus
1Würzburg University, Informatics Institute, Am Hubland, D-97074 Würburg, Germany
2Aachen University, Neurologic Clinic, Pauwelsstraße 30, D-52057 Aachen, Germany
Tutoring systems for medical education have become quite popular (see e.g. ). While many of them are based on the hypertext / hypermedia technique consisting of links between predefined windows, the idea of intelligent tutor systems (e.g. [12, 10]) is to generate the contents of the windows from underlying domain and didactic knowledge. For example, case oriented tutoring systems can be built in both ways: The patient case can be presented with a hypermedia system, where the sequence and the contents of the windows are prepared specifically for this case. Another approach is building a knowledge base capable of solving cases and using the correctly solved cases for tutorial purposes. Didactic knowledge is required for generating the incremental presentation of the case and for providing the user with feedback on his or her actions. While the costs for building hypermedia based training systems are directly proportional to the number of cases included, knowledge based training systems require a large initial effort to build and test the knowledge base and then only minimal costs for adding any number of new cases. The first system exploring this approach was GUIDON , a tutoring system on top of the expert system MYCIN. Insights gained from this work are, that a general problem solving method like backward chaining of rules in MYCIN severely restricts the explainability for tutorial purposes and that an unstructured rule format makes it difficult for the students to differentiate the key clause in the rule precondition from context and activation clauses.
While GUIDON needed additional knowledge for tutorial purposes, the general lesson is, that tutorial systems can be built on top of expert systems, but the requirements concerning the structure of the problem solving method and the contents of the knowledge base are considerably higher. Commercially available knowledge based training systems are the tutor versions of ILIAD  and QMR . They avoid the problems of GUIDON/MYCIN by a much simpler knowledge representation.
The general architecture of knowledge based tutorial systems is quite obvious: In addition to the basic components of expert systems - including knowledge base and problem solving, knowledge acquisition, explanation and interviewer component - specific tutorial features are case presentation and critic components; (see e.g. ). The main issues involved in designing such systems are practical evaluations regarding the influence of various problem solving methods, the sophistication of the knowledge representation, the case presentation technique depending on the amount of information to be presented, and the critic component depending on the users' time and motivation. In the following, the basic ideas for building and using a knowledge based neurology training system are described.
Developing a new training system with the diagnostic expert and tutor system shell box D3 [4, 11] only requires building a knowledge base and adding cases with the interviewer component. The knowledge base contains sufficient knowledge about the hierarchical or heterarchical structure of the findings to ensure automatic generation of (textual) case presentations in several modes, ranging from detailed presentations of findings to concentration on the key diagnostic elements for faster use. Early experiments  showed, that the basic practical problem is providing mechanisms for the user to maintain an overview on the vast amount of case data. Our solution is visualizing the heterarchical structure of findings in D3  with graphical hierarchies, where each node can be expanded or closed one level at a time by a mouse click similar to the user file selection in the Macintosh Finder. When starting a new case, the training system presents initial data about the patient. The user then selects tests to investigate his or her hypotheses. The system provides comments on demand for the users' actions and can also criticize his or her justifications. Critic of hypotheses is generated by comparing them with hypotheses inferred by the system from the same data the user has interpreted so far. Criticizing justifications is more difficult, because the system bases its conclusions on intermediate (pathophysiological) concepts derived from the raw data. Because the user is unaware of these intermediate concepts, they have to be compiled out for rating how much individual raw findings support hypotheses. The critic of the test choices is easily generated from the explicit representation of that knowledge in the knowledge base, where for each diagnosis a sequence of tests useful for its exploration is specified. When the user selects a test, the system compares it with its own choices (also considering second rate choices) based on the suspected diagnoses in the present stage of the tutorial session. The user can also justify test selection by reference to his or her suspected hypotheses, which the system criticizes with respect to the correctness of the suspected hypotheses and how well they can be investigated by the selected tests.
A large knowledge base has been built in the domain of neurology covering the diagnostic part of a standard textbook [Poeck 94] and is available with a voucher in that textbook. Building and testing the knowledge base took the author of the textbook two years with an average of about 1 to 2 hours per day. The knowledge base contains the main 120 diagnoses in neurology, about 1500 highly structured findings, about 300 intermediate concepts (finding abstractions) and more than 2500 rules. A typical session for an experienced user takes about 5 to 10 minutes per case. Currently (March 1995), more than 300 persons have requested the neurology trainer based on the voucher in the 9. edition of the textbook after four months of its publication and the delivery of the software has just started. Our plan for the next version of the neurology trainer includes a more sophisticated but still generated multimedial case presentation with graphics, fotos, videos, sounds and historical recordings. This gives the user the opportunity, not only to interprete findings but also to recognize them.
 Clancey, W.: Knowledge-Based Tutoring: The GUIDON-Program, MIT Press, 1987.
 Eysenbach, G.: Computer-Manual, Urban&Schwarzenberg, 1994.
 Fontaine, D., Beux, P., Riou, C. and Jacquelinet, C.: An Intelligent Computer-Assisted Instruction System for Clinical Case Teaching, Meth. Inform. Med. 33, 433-445, 1994.
 Gappa, U., Puppe, F., and Schewe, S.: Graphical Knowledge Acquisition for Medical Diagnostic Expert Systems, Artificial Intelligence in Medicine 5, 185-211, 1993.
 Lincoln, M. et al.: ILIAD Training enhances Students Diagnostic Skills. J. med. Systems 15, 93-110, 1991.
 Miller, R. and Masarie, F.: Use of the Quick Medical Reference (QMR) program as a tool for medical education, Meth. Inform. Med. 28, 340-345, 1989.
 Poeck, K.: Neurology, (in German), Springer, 9. Auflage, 1994.
 Poeck, K. und Tins, M.: An Intelligent Tutoring System for Classification Problem Solving, in Ohlbach, H.J.: GWAI-92: Advances in Artificial Intelligence, Springer, LNAI 671, 210-220. 1993.
 Puppe, B. and Puppe, F.: A Knowledge Representation Concept Facilitating Construction and Maintenance of Large Knowledge Bases, Methods of Information in Medicine 27, 10-16, 1988.  Puppe, F.: Intelligent Tutoring Systems (in German), Informatik-Spektrum 15, 195-207, 1992.
 Puppe, F., Poeck, K., Gappa, U., Bamberger, S., & Goos, K.: Reusable Components for a configurable Diagnostics Shell, (in German), Künstliche Intelligenz 2/1994, 13-18, 1994.
 Wenger, E.: Artificial Intelligence and Tutoring Systems, Morgan Kaufman, 1987.