Development of neural-network-based intelligent systems for medical pattern classification with missing features
Loading...
Date
2002-05
Authors
Mei Ming, Kuan
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This thesis is concerned with the development of neural-network-based intelligent
learning systems for pattern classi'fication problems with specific application to medical
diagnosis. The research is focused on the incremental learning ability of neural
networks in the presence of missing features in stationary as well as non;-stationary
environments. The incremental learning ability of the proposed intelligent systems is
adopted from the family of Adaptive Resonance Theory (ART) networks. In particular,
the Fuzzy ARTMAP (FAM) network has been selected as the basic network for the
development of the intelligent systems in this research.
In essence, many intelligent systems are applicable to idealistic environments and
circumstances in which complete information and data are available for learning.
Unfortunately, in many practical situations, it is not uncommon to encounter databases
that contain missing features that are caused by human factors as well as other
unavoidable factors, this often aggravates the complexity and diminishes the usability of
a learning system. The missing feature problem is an inevitable issue in real-world
databases, and 'it is therefore beneficial to devise strategies that are capable of
minimising the adverse affects of unexpected deviation and deterioration of
performance of intelligent learning systems caused by incomplete and/or deficient data.
In this research, the Fuzzy C-Means (FCM) clustering approach is embedded into the
FAM network in order to form a FAM-based intelligent system to handle the input
patterns with missing features. The proposed intelligent system is able t<? estimate the
values of missing features in the training samples, and to cluster the test samples with
missing features into distinct classes. As a result, the input patterns with missing
features can be employed for network training and, hence, all available information and
knowledge in the database can be fully utilised for devising the intelligent system.
A number of simulation studies using benchmark databases have been conducted to
evaluate the performance of the FAM-based intelligent system . .The performance of the
proposed system exhibits equivalent results when compared with other approaches. In
addition, real medical databases on the diagnosis of Myocardial Infarction (MI) and
Acute Coronary Syndrome (ACS) have been collected and experimented in order to
investigate the effectiveness of the proposed system in medical pattern classification
tasks. The results achieved indicate that the prediction accuracy could be improved
with increasing number of training samples, and demonstrate the potentials of the
proposed F AM-based intelligent system to learn continuously and autonomously using
databases with missing features in real environments.
Description
Keywords
Neural-network-based intelligent system , Medical pattern classification , Missing features