Lost in translation? Reliability of assessing inpatient antimicrobial appropriateness with use of computerized case vignettes.

Objective: To describe and measure reliability of a computer-assisted method of case vignette assembly and expert review to assess the appropriateness of antimicrobial therapy for hospitalized adults.

Design: Feasibility and reliability analysis of computer-assisted tool used to compare the effects of antimicrobial stewardship interventions.

Setting: Public teaching hospital.

Patients: Randomly selected adult antimicrobial recipients admitted to inpatient medicine services.

Methods: Clinical data abstracted from 504 paper medical records were merged with computerized laboratory and pharmacy data to assemble case vignettes that underwent expert review for appropriateness. We performed 3 validations, as follows: data for 35 vignettes abstracted independently by 2 research assistants were assessed for interrater agreement, expert review of 24 vignettes was compared with review of the corresponding paper medical records, and interrater reliability of antimicrobial appropriateness assessments by 2 experts was determined for 70 case vignettes.

Results: Vignette assembly and expert review each required 10-12 minutes per case. Potentially important discrepancies occurred in 0%-32% of clinical findings abstracted independently by 2 research assistants. Expert review of 24 vignettes and the corresponding full paper medical records yielded fair agreement (kappa, 0.30). The 2 experts identified inappropriate initial antimicrobial therapy in 67% and 61% of case vignettes reviewed independently; interrater agreement was improved after sequential case discussion and stringent application of appropriateness criteria (kappa, 0.72).

Conclusions: Our case vignette assembly and expert review method is efficient, but improvements in both technical and human performance are needed to be able to yield valid estimates of the prevalence of inappropriate antimicrobial use. Assessments of antimicrobial appropriateness require validation.