Background. Predicting antimicrobial resistance in gram-negative bacteria (GNB) could balance the need for administering appropriate empiric antibiotics while also minimizing the use of clinically unwarranted broad-spectrum agents. Our objective was to develop a practical prediction rule able to identify patients with GNB infection at low risk for resistance to piperacillin-tazobactam (PT), cefepime (CE), and meropenem (ME). Methods. The study included adult patients with sepsis or septic shock due to bloodstream infections caused by GNB admitted between 2008 and 2015 from Barnes-Jewish Hospital. We used multivariable logistic regression analyses to describe risk factors associated with resistance to the antibiotics of interest (PT, CE, and ME). Clinical decision trees were developed using the recursive partitioning algorithm CHAID (?2 Automatic Interaction Detection). Results. The study included 1618 consecutive patients. Prevalence rates for resistance to PT, CE, and ME were 28.6%, 21.8%, and 8.5%, respectively. Prior antibiotic use, nursing home residence, and transfer from an outside hospital were associated with resistance to all 3 antibiotics. Resistance to ME was specifically linked with infection attributed to Pseudomonas or Acinetobacter spp. Discrimination was similar for the multivariable logistic regression and CHAID tree models, with both being better for ME than for PT and CE. Recursive partitioning algorithms separated out 2 clusters with a low probability of ME resistance and 4 with a high probability of PT, CE, and ME resistance. Conclusions. With simple variables, clinical decision trees can be used to distinguish patients at low, intermediate, or high risk of resistance to PT, CE, and ME.
- antimicrobial resistance
- gram-negative bacteria