 |
Return to HomePage |  |
Return to Abstracts Index |
Medical diagnosis problems have three characteristics which usually makes difficult the applications of rule-based expert systems: the simultaneous presence of multiple disorders; the possibility of a disorder appearing without all of its manifestations; and the frequent modification of the knowledge base. These problems have been dealt with by conventional Artificial Intelligence (AI) approaches, with poor results, since, essentialy, they can be solved only by using the concept of learning, or adaptation In the present work, we have developed a medical-decision support system which combines knowledge representation using binary decision trees with the probabilistic approach, adding the possibility of modification of the tree by means of a learning procedure. The first time a medical knowledge base is created, it is empty. The physician must enter into the system all the manifestations present for a given patient, as well as the correct diagnosis. Henceforth, this first disorder becomes part of the current knowledge base, and can be diagnosed correctly from now on. Following the same process for other disorders and patients, the physician is able to build the decision tree by using some previously available database to train the system, in a specific medical area, such as cardiology or neurology. Larger and more comprehensive databases will lead to better results. Once a disorder or manifestation is entered by the user, it is kept by the system and it's not necessary to enter it again. Since the software knows a few manifestation he, it ask the user if the patient has it or not. Depending on the answer the computer will ask about another manifestation using a path based on the manifestations interrelations. These relations are dynamically calculated by the condicional probability. Using this, the system builds a rearrangeable net. With this structure, the the system can diagnose the partial manifestations cited before. The system needs to learn: 1. what's the root (the first manifestation to be asked); 2. how to run through the net; 3. when to stop asking and to choose a diagnosis; 4. associate subsets of manifestations with disorders.All of these items change as the system is used. Once the training period is finished, it can generate a tree-based expert system on the chosen area. The results obtained are compared to other systems
(Supported by a student fellowship to L.D.M. from FAEP/UNICAMP)
|
Return to HomePage | |
Return to Abstracts Index |