Department of Electrical Engineering and Computer Science
University of Wisconsin Milwaukee.
|
- BANTER - Bayesian Network Tutoring and Explanation BANTER provides a high-level user interface for Bayesian-network models, generates English-language explanations from probabilistic knowledge, and creates tutorial problems for instructional use. |
| Investigators |
|
| Background |
|
Bayesian networks have become the representation of choice for building
decision-making systems in domains characterized by uncertainty, and
have been applied to several medical domains. The models currently
available and under development provide a wealth of detailed knowledge
that can be used for educational purposes as well as clinical decision
support. Unfortunately, the information contained in these models is
not easily intelligible; tools are needed to make this information
comprehensible. The availability of shells for performing inferences
over Bayesian network models and the recent development of explanation
generation algorithms have made building such a tool possible.
|
| Current Status |
|
BANTER (Bayesian Network Tutoring and Explanation) is a
generic Bayesian-network shell that provides decision support and tutors
users in diagnosis and in selection of optimal diagnostic procedures.
BANTER can be used with any Bayesian network containing nodes that can
be classified as hypotheses, observations, and diagnostic procedures.
The system is designed so that the user need know nothing about Bayesian
networks in order to interact with it effectively. In fact, none of the
system's dialogs with the user indicates that the system is using a
Bayesian network to perform its reasoning. The user needs only some
knowledge of the particular domain and an elementary understanding of
probability.
BANTER computes the posterior probability of a diagnosis, determines the best diagnostic procedure to affirm ("rule in") or exclude ("rule out") a diagnosis, quizzes the user on the selection of optimal diagnostic procedures, and generates explanations of its reasoning. It can generate story problems and quiz the user on diagnoses and selection of optimal diagnostic procedures. Almost all of the system's reasoning is driven by the Bayesian network knowledge base; setting up the system for a new network requires minimal effort. BANTER transforms the information contained in a Bayesian network into an easily intelligible form for medical education and clinical decision support. BANTER quizzes and tutors users on the evaluation of diagnoses and optimal selection of diagnostic procedures. Since almost all the system's reasoning is performed using the Bayesian network knowledge base, configuring the system to work with a given network requires little effort. On the other hand, since nothing in the system's functionality indicates that it is using a Bayesian network for its reasoning, the complex details of the representation are hidden from the user. We currently are evaluating BANTER's explanatory content and style in tests with physicians at various levels of training. In addition to the model of gallbladder disease described above, we are applying BANTER to belief-network models for diagnosis of liver lesions by magnetic resonance imaging and echocardiographic diagnosis. |
| Future Efforts |
|
Future research will focus on (1) explanations in extremely large
networks, (2) more informative explanations, and (3) rigorous
evaluation. In the newly emerging models that contain thousands of
nodes, inference will become too slow to provide acceptable interaction
and the explanations produced by the current algorithm will become too
lengthy. For a given problem, typically only a portion of a given
network model will be relevant. We have developed a technique for
specifying a Bayesian network as a collection of rules in probability
logic and generating that portion of the network relevant to a given
computation. Integrating this technique into BANTER will significantly
reduce the complexity of inferences in very large networks.
BANTER provides more informative explanations by associating semantic information with Bayesian networks. Instead of displaying paths of influence with arrows, we indicate how each node influences its successor with terms like "causes" or "detects." Including abstraction information may make explanations more informative and more concise. Rather than explain only the current scenario, one could explain a more general scenario, of which the current one is an instance. For example, in the case of cholecystitis elevating temperature, one could additionally tell the user that any inflammatory disease, of which cholecystitis is an instance, has the tendency to elevate temperature. We plan to devise a World Wide Web interface for BANTER to make it platform-independent and to allow remote demonstration. |
| Software |
|
| Publications |
|