Automated classification of Pseudomonas aeruginosa from Gram-negative enterobacteriaceae.
Effective strategies to minimize the use of empirical broad spectrum antibiotics are necessary to decelerate the development of antimicrobial resistance (AMR) . However, the risk of infections by resistant bacteria such as Pseudomonas aeruginosa often compromises clinical decision-making. Although Gram staining is a quick and cost-effective procedure to identify causative pathogens, discerning Pseudomonas aeruginosa from more common Gram-negative enterobacteriaceae requires skilled personnel. To expedite Gram-staining-based decision-making on antibiotic treatment, here we developed a framework termed GramStainR for automated bacterial detection and classification from digital Gram stain images. Using reference isolates of Pseudomonas aeruginosa and Escherichia coli, the image-level classifier achieved the area under the curve (AUC) of 95%. A proof-of-concept, prospective cohort study using urine specimens obtained from patients with urinary tract infections (UTIs) yielded an image-level AUC of 82% (N = 96 images obtained from 10 specimens). Comparably high AUC (96%) was reproduced using a previously published external image dataset containing Bacteroides fragilis, Escherichia coli, and Psudomonas aeruginosa. These results suggest that automated image analysis is a promising aid for physicians selecting optimal antibiotic therapy before notified with culture results, thus warranting validation studies in diverse clinical contexts.