Abstract

Background: Despite its importance to global health, antibiotic resistance remains difficult to detect and quantify without extensive laboratory work. The clinical laboratory uses antimicrobial susceptibility testing (AST) to guide the treatment of infection, even though an accurate result can take several days to produce. We previously developed a flow cytometer-based AST method that rapidly generated data from antibiotic-exposed bacteria. Visualization and classification of complex multidimensional flow cytometer data by supervised machine learning presented a potential solution to the need for more timely antimicrobial therapeutic decision support. Method: We employed a three-stage machine learning process to visualize and classify the effects of antimicrobial exposure on bacteria in a series of flow cytometer experiments. Using data from our flow-assisted antimicrobial susceptibility test (FAST) method, we optimised each successive stage of analysis, to build an ensemble machine learning pipeline. We then compared the results generated by the ensemble-generated FAST result with proprietary flow cytometer software, and with the AST reference method (Minimum Inhibitory Concentration by Broth Microdilution). Finally, we applied the analytic pipeline to data from representative bacterial isolates from a clinical laboratory to assess the potential impact of the pipeline on the speed and accuracy of AST. Results: The three successive data machines (a) defined an antibiotic-unexposed population (AUP) of test bacteria, (b) classified the FAST result as within/outside antimicrobial concentration test range, and when within range (c) detected a concentration-dependent antimicrobial effect (CDE) from which we determined a predicted effective concentration (PEC) for subsequent comparison with the Minimum Inhibitory Concentration (MIC) broth microdilution (BMD) reference method. The complete pipeline generated an AUP, a susceptibility in-/outside tested range, a CDE, and a PEC. Reference culture collection Escherichia coli, Klebsiella pneumoniae and Staphylococcus aureus strains (ATCC 25922, ATCC 700603, ATCC 25923, ATCC 29213) tested against different antimicrobial agents (E. coli, K. pneumoniae: 13 agents, SEMPA1, SensititreTM, Oxoid, UK; and S. aureus: 16 agents, SEMSE3, Sensititreâ„¢, Oxoid) demonstrated a high degree of concordance between BMD MIC results and machine learning pipeline-analysed FAST results (categoric agreement of 91%, essential agreement of 100%). Detailed AST results were available from critical clinical isolates of E. coli and S. aureus on the same working day with only one major error (ME), the day before definitive BMD results. Conclusion: We used machine learning tools to visualise and classify antimicrobial susceptibility with the complex, multidimensional bacterial data sets generated by the FAST method. The speed at which antimicrobial susceptibility can be determined with the machine learning demonstrated here enabled same-day AST results from clinical isolates. This makes a case for using supervised machine learning to prototype automated or unsupervised analytical processes that improve the speed and throughput of rapid ASTs. When used by clinical microbiology services, the faster speed of AST result generation enabled by the combination of FAST and machine learning chould improve early antimicrobial therapeutic decisions, antimicrobial stewardship and identification of critical antimicrobial resistance. Funding: This project was unfunded, but data used to develop and test this pipeline were generated in projects funded by the Health Department of Western Australia, The Bill and Melinda Gates Foundation, ThermoFisher Scientific, the University of Western Australia and Lab Without Walls Inc., a locally-based not-for-profit organization, in which major equipment and small volume reagents were provided gratis by ThermoFisher Scientific, Division of Protein and Cell Analysis, Eugene, Oregon, USA. Declaration of Interest: Dr Inglis reports other from University of Western Australia, other from ThermoFisher Scientific, outside the submitted work. In addition, Dr Inglis and three other authors (TFP, KTM, CFC) have a patent PCT for flow cytometry-assisted susceptibility testing, managed by the University of Western Australia in association with the Health Department of Western Australia pending to University of Western Australia. Ethical Approval: No results obtained in this study were used in the clinical management of patients. Bacterial isolates from clinical specimens were obtained after sub-culture onto agar media in accordance with the Australian National Health and Medical Research Council guidelines on research ethics which advise that bacterial isolates from clinical samples can be used without patient consent or waiver.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call