Date of Award

August 2013

Degree Type

Thesis

Degree Name

Master of Science

Department

Health Care Informatics

First Advisor

Rohit Kate

Committee Members

Timothy Patrick, Ron Cisler

Keywords

Alcohol, Biomarkers, Recidivism

Abstract

Alcohol biomarkers, or naturally occurring molecules which occur in response to one's alcohol consumption, are proving to be a value tool in objectively monitoring one's alcohol consumption. Coupling this assessment tool, with advances in computing power, new and powerful predictions are becoming evermore possible. In this retrospective study, data was first collected that consisted of a sample of 249 drivers convicted of driving under the influence charge and who monitored over the course of a year by biomarker blood tests. This data was then analyzed using machine learning methods. TwoStep cluster analysis showed distinct drinking groups within the drivers who were monitored. In addition to this, a cost sensitive learning classifier was utilized in order to predict if a driver would relapse, having a subsequent driving under the influence arrest. The algorithm was able to predict 64% of the cases within the training set. Additionally, learning curves indicated that correctly classified cases increased with the increase of training data, indicating that predictions may become more accurate with the availability of more training data.

Download

Share

COinS