Association Between Removal of a Warning Against Cephalosporin Use in Patients With Penicillin Allergy and Antibiotic Prescribing

This cohort study assesses whether removal of a warning against use of cephalosporins in the electronic health record (EHR) of patients with penicillin allergy was associated with changes in the dispensing or administration of cephalosporins.


eMethods. Additional Methodology Details Study Population
Study inclusion and exclusion criteria were as follows: • Inclusion: o Members enrolled in KPSC (the intervention site) or KPNC (the comparison site) who received any oral or parenteral antibiotic (dispensed/administered) listed in eTable 1 between January 1, 2017 and December 31, 2018 • Exclusion: o Antibiotics dispensed/administered during the wash out period (the 7 days before and 7 days after the December 20, 2017 change in the alert) o Antibiotics dispensed/administered outside of a membership period o Patients with missing birth date or gender o Patients with birth date after the date of the first prescription, administration, or dispense (which suggested an inaccurate birth date). o Patients with death date before the first administration or dispense (which suggested an inaccurate death date).

Antibiotic Names and Categories
The list of 144 antibiotic generic and class names used to identify relevant medication and allergy records is shown in eTable 1.

Study Definitions
Antibiotic Use Periods of exposure to outpatient dispenses were calculated based on the dispense date and the days' supply, after adjusting sequential dispenses for stockpiled supply. Missing values of days' supply of outpatient medications were set to 1 day, and missing prescription dates were set to the date of the administration or dispense.
When two different antibiotics were administered in combination, they were considered to be two concurrent courses. For each course, we identified the date of the earliest prescription and the date of the earliest administration or dispense (the course start date). If any prescriptions in the course were associated with a diagnosis code, the earliest diagnosis code was used as the indication for the whole course.
For all courses, detailed route descriptions were categorized as either "oral" or "parenteral" to allow separate analyses by route of administration.

Anaphylaxis
We used ICD codes as described in the manuscript to identify potential cases of anaphylaxis. We included any outpatient coding of the relevant diagnoses on the same day as a parenteral course or within one day of an oral course of cephalosporin. For diagnoses that were coded during a hospitalization, it was not possible to determine from the secondary data the date on which the diagnosis occurred. Therefore, we included as potential cases all those coded during hospital or Emergency Department (ED) encounters that started within the same day as the start of a course of parenteral cephalosporin or within 1 day of a course of oral cephalosporin, or cases coded during a hospital stay that overlapped with the start of a course of a cephalosporin. Potential cases were confirmed by chart review, as described in the manuscript.

New Antibiotic Allergy
The calculations for new antibiotic allergies included courses for which the patient had at least 30 days of membership after the course start date, and included new allergy records and membership data through January 2019. The outcome was defined as a binary indicator for any new allergy records in the same antibiotic category during the 30-day follow-up window. New allergy records that appeared in the follow-up windows for multiple courses were assigned to the course(s) on the most recent date. If two courses in the same category started on the same date, the new allergy record was attributed to both (this accounted for approximately 2% of new allergy events).

Antibiotic Treatment Failure
The calculations for antibiotic treatment failure included courses of antibiotic monotherapy for which the patient had at least 30 days of membership in the analysis period after the course start date and included new courses of antibiotics through December 2018. A course was considered to be monotherapy if it was the only course of antibiotics that the patient started on the date. For this calculation, cephalosporin generations 1, 2, and 3 (or higher) were treated as different categories of antibiotics (see eTable 2 for the mapping of cephalosporin generic names to generations). If a new course of antibiotics started in the follow-up window of multiple courses, it was considered treatment failure only for the most recent course; and the outcome was a binary indicator for the presence of any treatment failure events in the follow-up window.

All-Cause Mortality
The all-cause mortality calculation included person-time starting at the earliest course start date in the period (to avoid immortal time bias) and ended at the earlier of the period end or the death date. Age was calculated as of the first course start date in the period, and penicillin allergy status was assessed as of the day before.

Hospital Days
The hospital days per person-year calculation included all periods of membership during the analysis period as person-time and included all hospital encounters during these member periods. Membership periods and hospital stays were separated into days in the pre and post periods, and days with and without a penicillin allergy record.

New Infections
The calculations for each new infection type (C. diff, MRSA, and VRE) were carried out separately. Person-time was calculated from the first course date in the period until the earlier of the end of the period or the end of the membership period. The date associated with a lab result was the specimen collection date. The date associated with an outpatient or ED diagnosis was the encounter date. The date associated with a hospital diagnosis with an inpatient admission was the encounter end date, unless the diagnosis was flagged as being present on admission (in which case it was associated with the encounter start date). Diagnoses and lab results from encounters in the three months before the start of the analysis period were included to determine whether infections in January-March 2017 were new. Because a person could have recurrent infections, the outcome was defined as a binary indicator of zero vs. one or more infections during the period.

Variables Used in Sensitivity Analyses
For the sensitivity analyses, dates of the onset of the individual comorbidities in the Charlson index, as well as hypertension and hyperlipidemia, were extracted based on ICD-10 diagnosis codes from hospital and outpatient encounters. We extracted Diagnosis-Related Group (DRG) data for hospital encounters that overlapped with course time periods, and used the DRG and Major Diagnostic Category (MDC) codes to identify courses that occurred during hospitalizations that included surgeries or labor and delivery. Encounters in Hospital Ambulatory Surgery (HAS) were also counted as surgery encounters.

Statistical Analyses
Summary A high-level summary of the outcome definitions and modeling strategies used in all main analyses is presented in eTable 3.
All models in the main analyses of the primary and secondary outcomes included the following variables: • We adjusted for race/ethnicity in the multivariate analyses because of known differences between the patient populations of the two sites. Race and ethnicity are self-reported as part of routine data capture within the EHR. We categorized the self-reported race/ethnicity values from the EHR into six categories (non-Hispanic American Indian/Alaska Native, non-Hispanic Asian/Pacific Islander, non-Hispanic Black, Hispanic, non-Hispanic White, and other). Other race/ethnicity includes those for whom race/ethnicity is unknown.

Modeling of the Primary Outcome
The primary outcome was change in antibiotic use, estimated at the course level. We used a generalized logistic regression model with penicillin as the reference level in the modeling of the primary outcome. The resulting estimate is an RROR representing the probability that a course of antibiotics will be cephalosporin.

Modeling of the Secondary Outcomes
We did not model probability of anaphylaxis because of small numbers (total of 68 potential cases, of which 9 were confirmed). To assess statistical significance of the anaphylaxis outcome (course level), we used Poisson regression on a person-level dataset with course count as an offset to calculate a chi-squared test with three degrees of freedom (based on two regions and two time periods).
We modeled all other secondary outcomes with binary logistic regression or Poisson regression, as described below.
We also completed sensitivity analyses that are described below; the study findings did not change based on these models.
• New antibiotic allergies: we used binary logistic regression models to calculate the RROR for a new allergy in the same antibiotic category within 30 days of the start of a course of antibiotics, combining all antibiotic categories. • Treatment failure: we used binary logistic regression models to calculate the RROR for the start of a course of antibiotics in a different category, within 30 days of the start of a course of monotherapy, combining all antibiotic categories. • Mortality: crude mortality rates were directly standardized to the age, gender, and race/ethnicity distribution of the full intervention site pre period population. Separately, we used Poisson regression to estimate the change in all-cause mortality rates as a RRRR. As a sensitivity analysis, we fit a second model that included indicators for the presence of comorbidities (described above) as of the first course start in the period. • Hospital Days: we used Poisson regression to estimate the change in hospital days as an RRRR. Because the same person could contribute both oral and parenteral courses to the analysis, this calculation could not be done separately for oral and parenteral courses. Instead, the model was also fit separately for patients who contributed parenteral courses to the analysis. As sensitivity analyses, we fit models that included indicators for the presence of comorbidities (described above) and with a more flexible negative binomial distribution. • New Infections: we used logistic regression with a complementary log-log link to calculate the change in new infection rates as RRRRs (C. diff, MRSA, and VRE). The models were also fit separately for patients who contributed parenteral courses to the analysis. As a sensitivity analysis, we recalculated the new infection outcomes without censoring at the first membership gap.

Modeling of Outcomes for Patients with and Without Penicillin Allergies
We fit additional models of the secondary outcomes to estimate the odds ratio (or rate ratio) of the outcome for patients with penicillin allergies vs. patients without penicillin allergies. The independent variables in these models were the same as in the main outcome models except that they did not include interactions involving the penicillin allergy indicator. That is, the models included region, period, the region-period interaction, penicillin allergy status, sex, age, race/ethnicity, and the age-sex interaction.

Calculation of Confidence Intervals
For all model results reported with confidence intervals, Generalized Estimating Equations (GEEs) were used to fit models with a random 20% subset of the patients to estimate a multiplicative correction factor for standard errors related to patient-level correlations (since patients often contributed multiple courses in the analysis). This correction factor was used to increase the size of 95% confidence intervals for the full model results. Running these models on 100% of the data was computationally infeasible. The correction factors for most outcomes ranged from 1.0-1.2. The exception was the hospital days outcome, where the correction factor ranged from 4.3-4.4. We considered the possibility that correlations at a higher level of aggregation might also affect standard error and confidence interval calculations. To explore this, we fit alternative models that included fixed effects for medical office building area.

Change in Antibiotic Use
Probability that a course of antibiotics will be cephalosporin.
Multinomial logistic regression to calculate an RROR.
Primary outcome.

Anaphylaxis
Counts of occurrence of anaphylaxis attributable to cephalosporin use, after a patient with a penicillin allergy uses a course of cephalosporin. Anaphylaxis must be coded on the same day as a parenteral course or within one day of the start of an oral course.
Not modeled due to small numbers.
This is the only outcome in the analysis that was limited to courses of cephalosporin among patients with penicillin allergies (this is the only outcome that involved manual chart review).

New Antibiotic Allergy
Probability of a new allergy record in the same antibiotic category, within 30 days of the start of a course of antibiotics.
Binomial logistic regression to calculate an RROR.
Events are assigned to the most recent course only.

Antibiotic Treatment Failure
Probability of the start of a course of antibiotics in a different category, within 30 days of the start of a course of monotherapy.
Binomial logistic regression to calculate an RROR.
Monotherapy is defined as being the only course of antibiotics that a patient starts on a date. Events are assigned to the most recent course only.

Intervention Comparison
Courses in the analysis 5,834,345 4,817,669 Courses in post period (%) 51.9 51.7 Courses that include any administrations (%) 20.1 21.9 Courses that include ED administrations (%) 3.9 6.6 Courses that include back-office administrations (%) 1.6 0.1 Courses that include any outpatient dispenses (%) 81.0 79.8 Courses that include a surgery encounter (%) 7.6 8.9 Courses that include a labor and delivery encounter (%) Other antibiotics (%) 3.5 3.3 6.0 3.8 3.8 3.6 5.6 5.3 The change in overall cephalosporin use was present in oral and parenteral courses separately, although the categories of antibiotics whose use decreased for patients in the intervention site with penicillin allergies differed slightly for the two routes. Both routes showed decreases in the use of clindamycin, macrolides, and quinolones. Oral sulfonamides, parenteral vancomycin, and parenteral "other antibiotics" also showed decreases.

eTable 8. Ratios of ratios of odds ratios for changes in antibiotic use for patients with penicillin allergies
Ratio of ratios of odds ratios Crude (from Table  2 In total, 10,475,367 courses (98% of all courses in the analysis) had 30 days of follow-up and were included in the new allergy calculations. We identified 61,566 courses with a new allergy record in the same category within 30 days, corresponding to an overall new allergy rate of 0.6%. New allergy rates v aried by antibiotic category and were more likely for patients that had a penicillin allergy, but generally decreased slightly in the post period. These general trends appeared in the data for all courses, as well as for oral and parenteral courses separately. The crude overall mortality rate was 18.0 per 1,000 person-years. The rate decreased slightly in the post period in both regions, and was higher for patients in the comparison site and patients with penicillin allergies.