Data Mining Antimicrobial Resistance: Hierarchical Clustering of Sensitivity Patterns across Thirty Antibiotics

Abstract

Introduction: Antimicrobial resistance is a major global public health threat, driven by inappropriate antibiotic use and associated with increased morbidity, mortality, and healthcare costs. In hospital settings, empirical antibiotic therapy is often initiated before susceptibility results are available, relying largely on conventional antibiograms. However, these provide only aggregated data and fail to reveal critical co-resistance patterns- information essential for predicting cross-resistance and optimizing treatment. To address this gap, our study applies hierarchical clustering to antimicrobial susceptibility data from a tertiary care hospital  to identify clinically relevant, multidimensional resistance relationships across antibiotics, with  the goal of  supporting evidence-based empirical therapy and strengthening antimicrobial stewardship efforts in resource-limited settings.

Methods: A retrospective observational study was conducted using microbiology laboratory data from a tertiary care hospital over a two years. Susceptibility results from clinical isolates were included. Hierarchical clustering using Ward’s linkage was applied to a Jaccard distance matrix to identify groups of antibiotics with similar resistance patterns, with analyses performed separately for each major bacterial species.

Results: Bacterial growth was detected in 31.3% (12,134/38,725) of clinical specimens processed over a two-year period. Predominantly gram-negative, the most frequent isolates were Escherichia coli and Klebsiella spp., with the majority originating from urine and respiratory samples. Species-specific co-resistance patterns were identified through hierarchical clustering. In E. coli, four antibiotic clusters emerged, distinguishing agents with retained activity from commonly used oral and fluoroquinolone antibiotics that demonstrated overlapping resistance. Klebsiella spp. exhibited clustering patterns consistent with multidrug-resistant phenotypes, in which reserve agents formed independent clusters.

Conclusion:  Hierarchical clustering revealed distinct, clinically relevant co-resistance patterns that highlight the heterogeneity in susceptibility and limitations of conventional antibiograms. Integrating this analytical approach into AMR surveillance can optimize empirical therapy, strengthen antimicrobial stewardship, and guide AMR control strategies in similar hospital settings.