More About LEXMED...


Learning Expert System for Medical Diagnosis, applied to Appendicitis

Why we chose Appendicitis:

Contrary to public opinion, the diagnosis of appendicitis is not simple. Surgeons it seems to be a fact that the clinical diagnosis of appendicitis is difficult and full of uncertainties, while avoiding an operation in case of an inflamed appendix means a high risk. On this background an average error rate of 15% or more (operations in case of not inflammed appendix) seem to express this problem. (Literature (german) see: B. Hontschik, 1994; vgl. Böhner et al. 1994). At the same time appendicitis is very common; (about 1 case per 1.000 inhabitants per year), the scientific literature is large, big databases of patients are available and could be used for testing. This makes a good background, to work on the improvement of the diagnosis with methods of probability theory and information theory.
 

Scientific background


Probabilities deliver a well researched language to express and conclude knowledge in domains with incomplete or uncertain information. For using probabilities in expert systems, some additional tasks had to be solved:
  1. However (as for every application and language) how to encode the knowledge of the domain within the given language
  2. how to reduce the degrees of freedom (which are always present given incomplete or uncertain knoweledge) in a fair way (i.e. for example not ad hoc)
  3. how to reduce the complexity of the calculations (necessary to derive the conclusions).
Our project is based on these problems and applies their solutions to medical diagnosis. We model medical knowledge by probability (frequency) statements ranging in intervals, we use the method of maximum entropy (see literature) to reduce the degrees of freedom and we use the properties of graphical models ('Independence Maps') to reduce the complexity of the calculations.
Essential to the project is therefore the use of the program system PIT , which has been developed at the Technical University of Munich to solve these problems. Its works with any set of (conditional) probability statements, ranging in intervals. In the case of our medical application we require the findings of the examination of the patient and ask the system for the probability for this patient to have a (certain) illness (e.g. 'perforated appendicitis').

This type of probabilistic expert systems could help the physicians in technical aspects (i.e. to obtain probable conclusions given some kind of knowledge). At the same time, it challenges the physician to be more explicit in his knowledge and open for discussion. We therefore hope that projects like this have the chance to improve the understanding and the quality of medical diagnosis.
 

Keywords

medical diagnosis, probability, incomplete knowledge, uncertain knowledge, artificial intelligence, method of maximal entropy, Data mining, machine learning, automatic classification, rules, scores, expert system, common sense reasoning,
Subjects:  mathematics, computer science, logics, medicine
 

Abstract (Flier) (German) .ps-file  .pdf-file


Pictures (Dr. Rampf, Dr. Ertel)