Incremental Data Mining for Active and Adaptive Knowledge Base for Patient Image Retrieval

Abstract

Introduction: The general perception that the use of information technology (IT) in health care is 10 to 15 years behind that in other industrial sectors such as banking, manufacturing and airline is rapidly changing. Faced with an unprecedented era of competition and managed care, health providers are now exploring the opportunities for using IT to improve quality while simultaneously reduce the cost of health care. Clinical decision support systems and expert systems (CDSSs / ESs) focus on utilizing artificial intelligence and data mining techniques to provide fast decision support for physicians. Although several success stories about CDSSs / ESs have been reported [Freudenheim 92, Nash 94], these systems usually lack the ability to adapt to pattern changes that are embedded in new data. This is due to the fact that the traditional algorithms utilized by these systems cannot learn on an incremental basis, i.e., once they are built, they cannot adjust their structures in which the knowledge is imbedded. Lack of incremental learning ability is not a unique phenomenon in health care expert systems. In fact, most of the machine learning algorithms developed to date are limited in their ability to adjust learned rules based on new, incoming data. In the Internet Age, when new data keep coming in at a high speed, this is a serious limitation for decision support systems. The main objective of this dissertation is to develop a new incremental neural network technique in order to support decision support systems' adaptive needs. An Incremental Neural Net (INN) algorithm that utilizes hidden layer activations to incrementally learn new patterns from incoming data is proposed. We then applied it to the Image Retrieval Expert System (IRES), a clinical decision support system for radiologists in University Medical Center (UMC), University of Arizona. The performance comparison between the INN and traditional neural net approach are compared. This chapter is organized as follows: section 1.1 briefly introduces the concept of data mining and incremental learning, which serve as technical foundations for this dissertation. Section 1.2 introduces the background of IRES project and describes its adaptive need. Section 1.3 addresses research motivation and objectives. Section 1.4 provides an overview of this dissertation.