Combining deep learning and droplet microfluidics for rapid and label-free antimicrobial susceptibility testing of colistin.

Efficient tools for rapid antibiotic susceptibility testing (AST) are crucial for appropriate use of antibiotics, especially colistin, which is now often considered a last resort therapy with extremely drug resistant Gram-negative bacteria. Here, we developed a rapid, easy and miniaturized colistin susceptibility assay based on microfluidics, which allows for culture and high-throughput analysis of bacterial samples. Specifically, a simple microfluidic platform that can easily be operated was designed to encapsulate bacteria in nanoliter droplets and perform a fast and automated bacterial growth detection in 2 h, using standardized samples. Direct bright-field imaging of compartmentalized samples proved to be a faster and more accurate detection method as compared to fluorescence-based analysis. A deep learning powered approach was implemented for the sensitive detection of the growth of several strains in droplets. The DropDeepL AST method (Droplet and Deep learning-based method for AST) developed here allowed the determination of the colistin susceptibility profiles of 21 fast-growing Enterobacterales (E. coli and K. pneumoniae), including clinical isolates with different resistance mechanisms, showing 100 % categorical agreement with the reference broth microdilution (BMD) method performed simultaneously. Direct AST of bacteria in urine samples on chip also provided accurate results in 2 h, without the need of complex sample preparation procedures. This method can easily be implemented in clinical microbiology laboratories, and has the potential to be adapted to a variety of antibiotics, especially for last-line antibiotics to optimize treatment of patients infected with multi-drug resistant strains.
Competing Interests: Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
(Copyright © 2024 The Authors. Published by Elsevier B.V. All rights reserved.)