Text

theses

A Bayesian network is a probabilistic graphical model that represents a set of variables and their conditional dependencies via one directed acyclic graph. For example, a Bayesian network could represent the probabilistic relationships between diseases and symptoms.
However, in some cases, the situation at hand does not lend itself to the single network model. Sometimes each observation represents a network, and so we are dealing with many networks rather than just one. We refer to these individual networks as precision networks. As an example, we may have a set of patients, each of which suffered multiple symptoms, conditions, and diseases referred to as events. These events may or may not be related to each other. A precision network, here called a precision disease network or PDN, may be created for each patient, and the total set of such PDNs can be stored and analyzed together.
In order to build such a PDN for each patient, we need to establish when events are related and when they are not. We developed a nonparametric algorithm that will determine whether such a relationship likely exists for two events, based on a data set with patients who experienced both. If such a relationship appears likely, we can provide an estimate of the proportion of dependent observations based on the time period between the two events. With the help of medical professionals, we may then establish an interval of time differences between those events within which we consider the events related, and outside of which we consider the events to be independent.
We note that medical researchers are often in need of finding new and interesting ideas for research within a topic. Those researchers will access the PubMed database and extract publications for the desired topic, usually resulting in a large amount of publications. They will then spend significant amounts of time perusing the abstracts of these publications in order to find an interesting idea that may be a candidate for a new clinical study.
We have developed a new method and computer application that examines all abstracts that fulfill the general search terms from bibliographic databases such as PubMed, mines those extracts for non-trivial, frequently occurring words, and allows for clustering of the abstracts using those words. By clustering and repeatedly re-clustering interesting clusters, a researcher can find an interesting subject for a new clinical study in a fraction of the time they spent previously.
We have also developed a new method to extract quality phrases from large volumes of text. Using this method, we have created an extension to the mining of abstracts that allows the clustering of quality phrases rather than words.

ETD_9603

Ph.D.

Includes bibliographical references

doi:10.7282/t3-hxm9-2f33

ETD doctoral

The author owns the copyright to this work.