759–4 Information Integration for Cardiology Decision Support

Abstract

Electronic medical records pose a challenge because of the complex types of data which are included. Decision support systems must be able to deal effectively with these data types. In the expert system demonstrated here, a diversity of data types are included. These data are processed by three different methods. However, the different methods of processing are transparent to the user. An overall rule-based interface integrates the different methods into one comprehensive system. Data types include crisp data, fuzzy data, temporal data, and numerical representation of chaotic analysis. Some data items which appear to be crisp, for example, test results, are more accurately represented as fuzzy numbers which indicate the degree of precision of the test. Four types of temporal data are considered: change in value from previous value, change in value relative to a specified time interval, duration data, and sequence data. A measure developed by the authors which determines the degree of variability in time series data is also included. The knowledge-based portion of the system utilizes approximate reasoning techniques which allows weighting of antecedents and partial presence of symptoms. The rule base is used as the interface which invokes a neural network model or time series analysis if certain rules are substantiated. The neural network model is a three-level feed-forward model based on a non-statistical learning supervised learning algorithm developed by the authors. Input data can be of any ordered form, including binary, categoric, integer, or continuous. The network can categorize data into two or more classes, and also produces a degree of membership for each class. Time series data, such as electrocardiograms, are important measurements for many diagnoses. An ECG may have an overall interpretation which can be used in the rule-based component, or categorized to be used in the neural network component. However, other analyses may also prove useful. In the application shown, a measure of variability for 24-hour Holter tapes is used. The combination of these techniques is illustration in a decision support system for the diagnosis and treatment of heart disease, including the use of a rule base, a supplementary neural network model of exercise testing data (ETT), and a time series analysis for Holter data.