Discussion

In this large EHR-based analysis of microbiology-tested older adults, AMR positivity and temporal patterns differed substantially by phenotype, care setting, specimen group, and area-level vulnerability. MRSA and ESBL had the highest phenotype-specific positivity; VRE and CRPA declined; and SVI associations were directionally heterogeneous. These findings favor phenotype-specific interpretation rather than a pooled AMR metric.

The observable Cosmos network expanded during the study. Tested volumes and contributing anonymous clinical locations increased for every phenotype, so annual estimates may reflect both within-location change and changes in the mix of locations, patients, settings, specimens, and testing practices. Facility and healthcare-system identities were masked, and clinical-location identifiers were not validated as one-to-one facility identifiers; a stable-facility cohort could therefore not be constructed reliably. Our models adjusted observed case mix and clustered by location, but these steps do not remove network-composition bias.

Our results address positivity among tested records rather than infection incidence and therefore differ from national burden studies. Jernigan et al estimated 622,390 multidrug-resistant infections among hospitalized US patients in 2017, with MRSA and ESBL-producing Enterobacterales accounting for 52% and 32%, respectively. ^{Reference Jernigan, Hatfield and Wolford16} The high MRSA and ESBL positivity in our tested population is directionally consistent, while the 43.2% clinical-specimen MRSA estimate also overlaps 2022 NHSN AR Option values. ¹⁷ CDC reported pandemic-era increases in several hospital-onset resistant infections, ¹⁸ whereas our mixed-setting older-adult population showed declining VRE and CRPA and stable MRSA. Differences may reflect setting, specimen and screening mix, operational definitions, stewardship and infection-prevention practices, and the evolving contributing network.

ESBL requires particular caution. Escherichia coli, Klebsiella pneumoniae, and Proteus mirabilis accounted for 89.8% of ESBL phenotype/proxy records, but third-generation-cephalosporin nonsusceptibility in other Enterobacterales may reflect AmpC or other mechanisms rather than confirmed ESBL production. ^11,12 The later ESBL increase is consistent with continuing US concern and Emerging Infections Program data showing predominantly E. coli and frequent community association. ^{Reference Duffy, Karlsson and Reses19} CRE remained uncommon after applying the corrected species-specific imipenem rule; the leading E. cloacae complex group may include ertapenem-only resistance from noncarbapenemase mechanisms, which could not be distinguished in the extract. ^{Reference Rankin, Stahl and Sabour20} CRAB estimates were less precise because of the smaller denominator and substantial overdispersion.

Higher SVI was associated with higher adjusted positivity for 5 phenotypes but lower ESBL positivity. Prior studies have linked socioeconomic disadvantage with community-associated MRSA, multidrug-resistant Enterobacterales, pneumococcal resistance, and resistant outpatient urine cultures. ^{Reference Chan, Westgard and Romasco21–Reference Mohanty, Ye, Sheets and Gupta25} The inverse ESBL association may reflect differential outpatient urine-culture density, community-onset exposure pathways, travel, ^{Reference Cars and Melhus26} and the composition of the phenotype/proxy rather than lower biological risk in more vulnerable communities. Race, ethnicity, payer, and SVI estimates remain social-context associations and should not be interpreted as biological or causal effects.

This study has limitations. First, the evolving EHR network and masked facility identities precluded a valid stable-facility analysis; temporal patterns may therefore reflect network composition as well as within-location change. Encounters, cultures, and outcomes outside participating Epic organizations were not observed. Second, the data did not consistently distinguish clinical infection from colonization or surveillance testing, and urine dominated the tested population. Third, ZIP-level 2020 SVI is less granular than census-tract linkage, may mask within-ZIP heterogeneity, and may not represent residence at each encounter. Fourth, operational phenotype definitions depended on available EHR interpretations and did not uniformly include MIC adjudication, enzyme confirmation, or carbapenemase testing. Fifth, models did not include patient or clinical-location fixed effects, CRAB was overdispersed, missing categorical values were modeled explicitly rather than imputed, and multiple comparisons were exploratory. Outcomes and treatment appropriateness were outside this study.

Strengths include the large, multi-year tested population; inclusion of multiple critical- and high-priority phenotypes; explicit phenotype-specific denominators; specimen, setting, and SVI context; and transparent reporting of annual tested volume and contributing locations. EHR-derived surveillance may complement facility antibiograms and established surveillance by describing tested populations across settings, but trends from an expanding network should be interpreted cautiously and not as national incidence or population burden.