Show simple item record

Item Description Kulikowski, Casimir Alexander en_US 2009-09-09T19:36:44Z en_US 2009-09-09T19:36:44Z en_US 1970 en_US
dc.identifier.uri en_US
dc.description Typescript. en_US
dc.description Thesis (Ph. D.)--University of Hawaii,1970. en_US
dc.description Bibliography: leaves [167]-169. en_US
dc.description ix, 169 l illus., tables en_US
dc.description.abstract This dissertation describes a pattern recognition model which has been successfully used to simulate a doctor's diagnostic process. A computer program implementing this model can be a valuable aid to the specialist, freeing him from routine screening procedures and making his skills more readily available to those patients who need them. The process of diagnostic inference is formulated as a pattern recognition task for which each disease category is represented by a characteristic pattern of symptoms and other patient variables. The method of class featuring information compression that was used assumes these characteristic patterns to form a subspace of the variable space. The subspaces are defined in terms of the components of an optimal expansion for the data vectors of a class. The use of this optimal expansion guarantees that the representative samples from which it is calculated will lie closer, on the average, to a subspace spanned by its own principal components than to any other subspace of the same dimension. The principal results of this dissertation in the pattern recognition field are procedures for selecting the dimensionality of the class subspaces to obtain good discrimination. Two quantities were found to be good predictors of recognition performance. One is a ratio of the inclusions of the paradigms of two classes within the subspace of one of them; the other is the average margin of correct classification for the paradigms of a class. Both, under certain conditions, are highly correlated with recognition performance. Therefore, a procedure for subspace selection is the maximization of the average margin of correct classification for one class subject to a constraint on the margins for all other classes. A similar procedure can be derived with the ratios. The average inclusions of paradigms within a subspace can be calculated from the autocorrelation and projection matrices of the classes. Thus, the ratios and margins are found and performance predicted without the need of actually performing any classifications. The model was tested with data obtained for 3291 patients who were examined at the Straub Clinic in Honolulu between 1963 and 1969 for the possibility of thyroid dysfunction. Data from 1963 to 1968 were used as paradigms and 1969 data as a test sample. In the diagnosis of hyperthyroidism the pattern recognition program performed consistently better than a linear discriminant method. In the diagnosis of hypothyroidism, however, the original class featuring information compression program did not perform as well as the other methods with which it was compared. Subspace selection by the method of constrained margin maximization improved the performance of the pattern recognition program considerably. The results indicate that this method can serve as a good representation of a realistic diagnostic situation. In order that the method be useful clinically a sequential version of the program was developed giving the classification of a patient at every stage of diagnosis. A category of deferred judgment was included in order that more data could be gathered when a diagnosis was uncertain. The sequential program satisfied clinical tolerances of accuracy as determined by a specialist. The on-line performance of this program has been simulated successfully and is to be implemented clinically. en_US
dc.language.iso en-US en_US
dc.publisher [Honolulu] en_US
dc.relation Theses for the degree of Doctor of Philosophy (University of Hawaii (Honolulu)). Electrical Engineering; no. 313 en_US
dc.rights All UHM dissertations and theses are protected by copyright. They may be viewed from this source for any purpose, but reproduction or distribution in any format is prohibited without written permission from the copyright owner. en_US
dc.subject Electronic data processing -- Medicine en_US
dc.subject Diagnosis en_US
dc.title A pattern recognition approach to computer-aided medical diagnosis en_US
dc.type Thesis en_US
dc.type.dcmi Text en_US

Item File(s)

Description Files Size Format View
Restricted for viewing only uhm_phd_7112220_r.pdf 3.956Mb PDF View/Open
For UH users only uhm_phd_7112220_uh.pdf 3.906Mb PDF View/Open

This item appears in the following Collection(s)

Show simple item record


Advanced Search


My Account