Progressing From Association to Causation: Examining Prior Antibiotic Exposure and Antimicrobial Resistance Across Three US Health Systems Using Causal Machine Learning.

BACKGROUND: Antibiotic exposure is hypothesized to drive antimicrobial resistance, but confounding by indication has limited causal inference from observational data. We estimated the effect of prior exposure on subsequent same-class nonsusceptibility across 7 antibiotic classes and 3 US health systems using a causal-inference framework. METHODS: We analyzed 1 165 151 culture episodes from Mass General Brigham (MGB), Stanford Health Care, and Beth Israel Deaconess Medical Center (BIDMC; via MIMIC-IV). The causal effect of 90-day prior exposure on same-class resistance was estimated by double machine learning, adjusting for more than 35 confounders per site, including prior same-class resistant isolates, infecting-organism group, and calendar period. Robustness was assessed via inverse probability weighting, E-values, exposure-window and dose-response sensitivity, and negative-control pairs. Rates of microbiologically inactive empiric therapy were compared by prior same-class exposure. RESULTS: Prior fluoroquinolone exposure was associated with increased nonsusceptibility by 11.9 and 12.6 percentage points at the primary sites (MGB, Stanford) and 8.6 points in validation (BIDMC) (all P < .001). Positive associations occurred for all 7 classes at all sites (sulfonamides indeterminate at Stanford), strongest at 30 days and attenuating through 365 days. Among patients with prior fluoroquinolone exposure receiving empiric fluoroquinolones for bloodstream infection, therapy was inactive in 50%-71% of episodes versus 27%-37% without prior exposure. Across classes, 3309 inactive empiric prescriptions were associated with prior same-class exposure. Median E-value was 2.5. CONCLUSIONS: Prior antibiotic exposure was consistently associated with increased same-class nonsusceptibility across antibiotic classes and health systems, paralleling higher rates of microbiologically inactive empiric therapy. These findings support prospective evaluation of antibiotic-history-informed empiric prescribing.