[New England
      Complex Systems Institute]
[Home] [Research] [Education] [Current Section: Activities & Events] [Community] [News] [The Complex World] [About Complex Systems] [About NECSI]
International Conference on Complex Systems (ICCS2006)

Communicable disease outbreak detection and emergence of etiologic phenomena in an evolutionary computation system

John Holmes
University of Pennsylvania School of Medicine/Faculty

     Full text: Not available
     Last modified: April 25, 2006

Abstract
Disease outbreaks are complex events that are evolutionary in nature. They donít occur all at once, but rather over time often in dynamical patterns. These patterns are mediated by many factors such as the underlying complexity of human (and in the case of zoonoses, animal-human) social networks, characteristics of the underlying etiologic agent, individual immunity and person-based immune variation, population-based (herd) immunity, and environmental characteristics such as climate, prevailing wind patterns, and building characteristics. These etiologic phenomena are important to the emergence of a disease outbreak, but even more so is the complex interaction between them. Current approaches to outbreak detection focus on a signal occurring over some time period, usually within a given geographical area. Signals include the appearance of syndromes or purchases of over-the-counter medications as proxies for detecting an underlying disease. As essential as these types of systems are, they usually ignore etiologic phenomena and their interactions that may help to distinguish between true and false positive outbreak signals. This paper reports on a method for detecting such phenomena using an evolutionary computation paradigm, the accuracy-based learning classifier system, EpiXCS. EpiXCS uses a knowledge base consisting of a fixed population of individual classifiers in a genotype-phenotype representation that translates directly to condition-action rules. Reinforcement learning is used to reward accurate classifiers in Markov environments, such as one finds in outbreak events, and a genetic algorithm is used to search the solution space and to propose plausible new classifiers. When applied to several simulated and real outbreak datasets, EpiXCS detects etiologic factors missed by traditional methods such as Bayesian classification and logistic regression. With its support for rule visualization EpiXCS should prove to be a valuable tool for modeling simulated outbreaks as well as detection of actual outbreaks in real-time.




Conference Home   |   Conference Topics   |   Application to Attend
Submit Abstract/Paper   |   Accommodation and Travel   |   Information for Participants


Maintained by NECSI Webmaster    Copyright © 2000-2005 New England Complex Systems Institute. All rights reserved.