Diagnostic Reasoning with Surface and Deep Level Knowledge in Medical Domain

Abstract

The aim of this work is to structure the domain knowledge based upon some conceptual models so that the knowledge can be efficiently used for diagnostic reasoning in medical domain. Medical domain is complex and incompletely specified, hence the unstructured Al techniques, normally used for engineered systems, become inefficient. We have developed and implemented two expert systems based upon the models for diagnostic reasoning using surface and deep level knowledge. Organizing the associational knowledge, between manifestations and disease categories, around a single taxonomy is a useful idea. Inorder to organize the medical knowledge in a single taxonomy, it may be necessary to introduce artificial categories and other simplifying assumptions. In this work, a model has been proposed that organizes the disease category knowledge around multiple taxonomies. Here, each taxonomy specializes in classifying a disease category based upon either a disease state of an organ, or the class/subclass relationships among disease causing organisms, or the structural relationships among various organs. A disease category can belong to more than one taxonomy and in each taxonomy it plays the role of a classifier specializing in the type of taxonomy. Each category is represented as a summary of all the roles it plays in various taxonomies and as having two aspects. The situational aspect distinguishes a category from others in different taxonomies using the manifestations it exhibits. The control aspect describes how a category, depending upon the local factors, can be specialized or generalized in the taxonomies to which it belongs. Each disease category is represented as a mixture of production rules and frames. These ideas have been implemented in a system, INTERX, in UCI-LISP on a DEC-1090 system for the domain of infectious diseases. The second system diagnoses diseases using the deep-level pathophysiological knowledge of the disease mechanisms. The system uses Qualitative Simulation (QS) for generating the abnormal behavior of human body. Since the natural systems, e.g. human body, are too complex and incompletely understood by the domain specialists, the techniques developed for engineered systems make QS inefficient. We developed several techniques to make QS efficient, (i) The disease is a combined behavior of various organ systems reacting to a fault arising in the body. Since this knowledge is vast and the organ systems react differently to a disease, an organization of the domain knowledge into subsystems at different levels of abstractions called Local subsystem(LOC) and Environment (ENV), is proposed. The fault considered here is the entry of a foreign organism into the body. The LOG depicts the disease mechanisms that are directly caused by the disease causing organism whereas ENV depicts their side effects, (ii) The disease mechanisms are described as processes that can bring state changes within the organs. Each process has temporal, structural, causal and effect aspects. Since the knowledge available of a process is incomplete, the temporal knowledge is represented as ambiguous intervals. (iii)These processes vary in their abstractions depending upon the subsystem to which they belong. The LOG processes are described using the substances and structural locations to which they belong. The ENV processes are described using variations of systemic parameters. The behavior of complex system is generated through a predictive analysis, using processes, and is represented as causal networks of disease states of structural locations. During the behaviour generation, assumptions are generated to resolve the ambiguities reulting from the incompletely specified processes. The predictive analysis predicts the manifestations to verify the hypothesis about the organism entry into body. The causal networks are generated, abstract ones, initially, for large classes of organisms, and more detailed ones for specific organism types. The abstract causal networks control the generation of alternate simulations in the more specific networks through an Assumption Inheritence mechanism. The control knowledge of the system implemented, DIAN, is represented as a hierarchy of operators which interact through a message passing mechanism. The system, DIAN, is implemented in UCI-LISP on a DEG-1090 system for respiratory tract Infectious diseases.