Antilogic, a new supervised machine learning software for the automatic interpretation of antibiotic susceptibility testing in clinical microbiology: proof-of-concept on three frequently isolated bacterial species

Abstract

ObjectiveAntibiotic susceptibility testing (AST) is necessary in order to adjust empirical antibiotic treatment, but the interpretation of results requires experience and knowledge. We have developed a machine learning software that is capable of reading AST images without any human intervention and that automatically interprets the AST, based on a database of antibiograms that have been clinically validated with European Committee on Antimicrobial Susceptibility Testing rules. MethodsWe built a database of antibiograms that were labelled by senior microbiologists for three species: Escherichia coli, Klebsiella pneumoniae, and Staphylococcus aureus. We then developed Antilogic, a Python software based on an original image segmentation module and supervised learning models that we trained against the database. Finally, we blind tested Antilogic against a validation set of 5100 photos of antibiograms. ResultsWe trained Antilogic against a database of 18072 pictures of antibiograms. Overall agreement against the validation set reached 97% (16 855/17 281) regarding phenotypes. The severity rate of errors was also evaluated: 1.66% (287/17 281) were major errors and 0.80% (136/17 281) were very major errors. After implementation of uncertainty quantifications, the rate of errors decreased to 0.80% (114/13 451) and 0.42% (51/13 451) for major and very major errors respectively. DiscussionAntilogic is the first machine learning software that has been developed for AST interpretation. It is based on a novel approach that differs from the typical diameter measurement and expert system approach. Antilogic is a proof of concept that artificial intelligence can contribute to faster and easier diagnostic methods in the field of clinical microbiology.