Neural networks for predicting antibiotic resistance - an analysis of performance

Abstract

This thesis investigate the possibility of predicting bacteria resistant to antibiotic treatment. The investigation is a part of developing a diagnostic tool for hospitals to use when deciding which antibiotic a patient should be treated with. This tool will be more and more relevant as antibiotic resistance and multidrug-resistance is spreading in the world with an accelerating rate. Neural networks are used for predicting the probability of a bacterium being resistant to an antibiotic. The architecture of these networks is investigated and their outcome and performance are analysed. The distribution of the predictions are investigated and then different classification limits are tested. A classification limit gives a number of unclassified samples and errors, and these samples are investigated with regards to age, gender and country. Finally, the performance of the model is measured as the number of tested antibiotics is reduced. When investigating the architecture of the neural networks the result are quite similar regardless to the number of layers and neurons. For the predictions, some antibiotics are more easily separated than other. This leads to a big variation in the number of unclassified samples and error between different antibiotics. When analysing the unclassified samples and errors only country could affect the predictions. The performance of the model when having more than four antibiotics tested as input is high. The conclusion is therefore that predicting antibiotic resistance using neural networks is possible and could potentially be used to replace measurements in the hospital laboratory.