DATA MINING FOR HEALTHCARE DECISION SUPPORT

点击翻译按钮获取中文摘要

The broad, long-term objective of this research project is to help improve the quality and cost-effectiveness of health care delivery through the use of innovative mathematical modeling, design of efficient and novel solution approaches, and effective use of the powerful new computing technologies that can facilitate knowledge discovery in large databases. While methods drawn from computer science and statistics disciplines have traditionally been used for this problem, there are significant opportunities for exploring the capabilities of combining methods drawn from the operations research discipline with the traditional techniques, such as using linear and nonlinear programming algorithms for improving neutral network design and performance, and metaheuristic search for improving genetic algorithm design and performance. Drawing on these recent developments, this project aims to develop a hybrid of computer science and operations research based methods of mining large databases. The potential of this new class of methods will be demonstrated using a high quality clinical database developed at the University of Pittsburgh Medical Center through funding from the Agency for Health Care Policy and Research. This database contains extensive information on patients with community-acquired pneumonia (CAP). The specific focus of this project is to address the problem of predicting patient mortality in the area of CAP. Pneumonia is an important problem to investigate because it affects a significant group of people, leads to complications requiring expensive hospitalizations, and is the sixth leading cause of death in the US. This study proposes to predict mortality of hospitalized patients based on findings recorded during the initial patient-physician encounter using a prediction technique known as probabilistic belief networks. This method will be extended by combining the special features of a metaheuristic strategy called tabu search to help reduce the complexity of the search process during the design and construction of probabilistic belief networks. Additional features of tabu search, such as scatter search, path relinking, and probabilistic tabu search, facilitate the efficient exploration of the search space for learning belief networks from data. This new class of methods will be tested on real data from a large clinical database on pneumonia, and compared with the capabilities of a number of machine learning methods that have previously been applied to the same problem.