Authors:
M. L. Tlachac
1
;
Elke Rundensteiner
1
;
Kerri Barton
2
;
Scott Troppy
2
;
Kirthana Beaulac
3
and
Shira Doron
3
Affiliations:
1
Worcester Polytechnic Institute (WPI), United States
;
2
Massachusetts Department of Public Health (MDPH), United States
;
3
Tufts Medical Center, United States
Keyword(s):
Antimicrobial Resistance, Antibiotic Resistant Bacteria, Antibiograms, Predictive Analytics, Regression, Support Vector Regression, Model Selection.
Related
Ontology
Subjects/Areas/Topics:
Artificial Intelligence
;
Biomedical Engineering
;
Business Analytics
;
Cardiovascular Technologies
;
Computing and Telecommunications in Cardiology
;
Data Engineering
;
Decision Support Systems
;
Decision Support Systems, Remote Data Analysis
;
Health Engineering and Technology Applications
;
Health Information Systems
;
Knowledge-Based Systems
;
Pattern Recognition and Machine Learning
;
Symbolic Systems
Abstract:
Antibiotic resistance evolves alarmingly quickly, requiring constant reevaluation of resistance patterns to guide empiric treatment of bacterial infections. Aggregate antimicrobial susceptibility reports, called antibiograms, are critical for evaluating the likelihood of effectiveness of antibiotics prior to the availability of patient specific laboratory data. Our objective is to analyze the ability of the methods to predict antimicrobial susceptibility. This research utilizes Massachusetts statewide antibiogram data, a rich dataset composed of average percent susceptibilities of 10 species of bacteria to a variety of antibiotics collected by the Massachusetts Department of Public Health from over 50 acute-care hospitals from 2002 to 2015. First, we improved data quality by implementing data filtering strategies. We then predicted up to three future years of antibiotic susceptibilities using regression-based strategies on nine previous years of data. We discovered the same predi
ction methodology should not be utilized uniformly for all 239 antibiotic-bacteria pairs. Thus, we propose model selection strategies that automatically select a suitable model for each antibiotic-bacteria pair based on minimizing those models' mean squared error and previous year's prediction error. By comparing the predictions against the actual mean susceptibility, our experimental analysis revealed that the model selectors based on the predictions of the previous performed best.
(More)