Statistical Characterization of Antimicrobial Activity: Do Common Descriptions Hold Up To Unsupervised Data Analytics?
Background: Antibiotic spectrum is commonly described as it relates to structure or activity against resistant pathogens. Data analytics is a method of verifying these descriptors and potentially describe new patterns. Principal components analysis (PCA) is a data analytic technique that produces new uncorrelated variables in order of decreasing variability. This project examines principal components (PC) of antimicrobials as they relate to spectrum of activity.
Methods: Previously published data were utilized to determine antimicrobial activity of 70 antibiotics against 55 bacterial pathogens. Antibiotics were considered active, partially active or inactive against bacteria. Dimensionality was reduced via principal components analysis. PC1 and PC2 were correlated with clinically relevant groups of bacteria.
Results: Doripenem, imipenem, and ticarcillin-clavulanic acid had the largest PC1 scores (2.65, 2.62, and 2.53, respectively), while, daptomycin, teicoplanin, and telavancin had the smallest PC1 scores (-2.48, -2.40, -2.38, respectively). Tigecycline, chloramphenicol, and ampicillin-sulbactam had the highest PC2 scores (1.84, 1.74, and 1.50, respectively). While, tobramycin, amikacin and gentamicin had the lowest PC2 scores (-1.99, -1.99, and -1.97). PC1 score correlated to number of Gram-negative bacteria (R2 = 0.942, p < 0.001, Figure 1). PC2 score correlated to number of non-Gram-negative bacteria (R2 = 0.858, P < 0.001, Figure 2).
Conclusion: PCA suggested Gram-negative coverage was the major source of antimicrobial variability. PC2 described a rich interplay of Gram-positive, anaerobic and atypical activity. PCA is a valuable tool for comparing spectrum of activity between antibiotics. PC scores may also be useful in future studies on antimicrobial de-escalation.
Figure 1. Principal Component 1 Score vs. Gram-Negative Activity
Figure 2. Principal Component 2 Score vs. Non-Gram-Negative Activity
M. Garfinkel, None
J. Gervasio, None
K. Lyndaker, None
C. Bergeron, None