Predicting Antibiotic Resistance Phenotypes using Neural Networks
Date
Authors
Type
Examensarbete för masterexamen
Model builders
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Antibiotic resistance is one of the biggest threats to human health. Today it causes
700 000 deaths per year, a number that is estimated to rise to 10 million by 2050.
Measuring antibiotic resistance is time consuming and while waiting patients may
receive ineffective treatment. It is thus interesting to see what previous measurements
can tell us about antibiotics that have not been measured. In this thesis
neural network based imputation is investigated as a suitable method for prediction
of antibiotic resistant phenotypes. This was done using comprehensive data
collected by the EARS-Net. The data contains bacteria from 330000 patients that
have been tested for resistance against a range of different antibiotics. In this thesis
neural networks were trained to predict missing resistant and susceptible phenotypes
based on available resistance data. In addition the networks had access to
information about the pathogen and the reporting country. After tuning the number
of nodes and training the networks on 85% of the available data we received an
average error rate of 5% when testing on the remaining data. The average very major
error rate, i.e. the error rate on measurements whose true value were resistant,
was 11.8%. The error rates varied for different antibiotics and some had very major
error rates below 5%. Neural network based imputation shows promise as a method
for predicting antibiotic phenotypes and with further research into what affects the
neural networks performance it could be a useful tool when measuring antibiotic
resistance.
Description
Keywords
Antibiotic Resistance, Neural Network, Imputation, Categorical Variables.