Network-based characterization of latent co-testing patterns in antimicrobial susceptibility testing

Antimicrobial resistance (AMR) surveillance depends on antimicrobial susceptibility testing, yet antibiotic panels vary by pathogen cohort, specimen, year, ward, and care setting. We term these empirically reconstructed testing architectures shadow antibiograms: structures that shape which susceptibility results enter surveillance datasets and which resistance relationships can be observed. We analyzed approximately 13 million bacterial isolates from Germany's National Antibiotic Resistance Surveillance system (2019-2023) across six WHO BPPL-aligned operational cohorts. Co-testing networks were constructed using Jaccard, Dice, cosine, and phi-coefficient similarity, evaluated with Fisher's exact tests and Benjamini-Hochberg correction, and partitioned with multi-resolution Louvain community detection. Networks contained three to four dominant co-testing communities; false discovery rate (FDR) retention ranged from 89 to 100% across cohorts and specimen types. Jaccard provided the best balance of coherence, stability, antimicrobial-class alignment, interpretability, and edge retention. This work provides one of the first national-scale empirical characterizations of shadow antibiogram structure, supporting diagnostic stewardship and bias-aware AMR surveillance interpretation.