Study design
This was a retrospective cohort study of electronic medical record (EMR) data from an IDN, from 1 July 2016 to 31 March 2020 (Fig. 1). Data were broadly representative of treatment patterns in Virginia, Maryland, and Washington DC. The database contains information on patient diagnosis, prescriptions, procedures, and laboratory values, inpatient visits, and outpatient visits among others. We utilized data from one of the largest IDNs in the Mid-Atlantic region of the US, serving more than 5 million patients.
Antibiotic resistance was reported through antimicrobial susceptibility tests of the isolate in the EMR, rather than at the patient level. The number of isolates was determined and then a decision rule was used to assign a person to a susceptible/not-susceptible status. Sensitive results were categorized as susceptible, whereas resistant and intermediate results were categorized as not-susceptible. For patients with two or more isolates with different susceptibility results (i.e., sensitive results for one isolate and resistant results for another isolate), the individual was classified as being not-susceptible at the person level.
To identify oral antibiotics which are indicated for uUTI, we used the IDSA 2011 guidelines [3] and data published by Brusch et al. [10].
Patients
Eligible patients were females aged ≥ 12 years, diagnosed with a UTI based on International Classification of Disease, 9th Revision (ICD-9) and ICD-10 coding in the primary or secondary positions, i.e., patients who had ≥ 1 urine culture with 10,000–100,000 colony-forming units (CFU)/mL, had ≥ 1 test to determine antibiotic susceptibility, and received oral antibiotics within ± 5 days of the index uUTI diagnosis to ensure patients had a symptomatic UTI and were receiving empiric therapy with an antibiotic. This treatment window was designed to capture patients with varying periods between diagnosis and treatment and was not designed to represent a change in therapy based on the availability of susceptibility results.
Exclusion criteria were used to further refine the study population to ensure that cases related to cUTI were excluded; a clinical expert was consulted to determine the appropriate exclusion criteria. Individuals were excluded if they were pregnant; had human immunodeficiency virus/acquired immunodeficiency syndrome (HIV/AIDS) with antibiotic use from 6 months to 6 days before the index uUTI date; had a urinary catheter at index UTI event or within 48 h; had symptoms of systemic illness, such as fever (temperature ≥ 38.3 °C), nausea or vomiting, reported or reproduced flank pain at index UTI event or within 48 h of the index UTI event; received intravenous (IV) antibiotics as initial therapy; had renal or urologic abnormalities, immunocompromising conditions, or chronic conditions that lead to systemic infections. Patients with missing laboratory values or utilization measures were also excluded.
Objectives
The primary objective was to measure the difference in healthcare resource use and associated costs (both all-cause and UTI-related) between patients with susceptible and not-susceptible urine isolates in the 6 months post-index date (inclusive). The primary outcome was related to the EMR encounter data (e.g., hospitalizations, office visits, emergency department [ED] visits, etc.) and the imputed cost associated with each of these.
Secondary objectives included measuring the difference in medical, pharmacy, and antibiotic costs (all-cause and UTI-related) in the 6 months after uUTI between patients with uUTI isolates that were susceptible and not-susceptible. In addition, we assessed the differences in hospital admissions, ED visits, urgent care visits, office visits, and laboratory tests in the 6 months after the index uUTI diagnosis. We also evaluated the difference in probability of uUTI progressing to cUTI between patients who had susceptible or not-susceptible isolates. In terms of disease progression, cUTI was defined as a combination of new fever, nausea, or vomiting, in addition to uUTI symptoms; or receipt of an IV antibiotic 3–28 days after index uUTI. The index date was defined as the first uUTI diagnosis/urine culture with a prescription for an antibiotic with no prior uUTI in the last 28 days.
Since information on prescription fills within the IDN system would be incomplete if the prescription was filled outside the IDN, our analysis relied on medication prescriptions rather than actual use.
The key dependent variables for the analyses were difference in cost and healthcare utilization and the key independent variable was whether an infection involved isolates that were susceptible or not-susceptible. Several other independent variables were included in the analysis as covariates to inform the propensity score matching. These included demographics (age, race, ethnicity), health insurance plan, and Charlson comorbidities (hemiparesis, renal disease, myocardial infarction, chronic pulmonary disease, rheumatic disease).
Cost calculations
Medical costs were calculated using Medicare fee-for-service rates. However, these rates were then adjusted for payer mix of our data using the relative reimbursement for inpatient stays across payers as reported based on hospital charge and cost-to-charge ratios identified in the Healthcare Cost and Utilization Project [11]. For inpatient costs, we used 2018 100% Medicare limited dataset inpatient and calculated adjusted total paid to 2020 quarter (Q) 1 dollars (using Consumer Price Index for Medical Care) for each diagnosis-related group program. The process for mapping outpatient cost was more heterogeneous. We first calculated the adjusted total paid per Current Procedural Terminology (CPT) code from 5% Medicare outpatient and carrier files. We primarily used outpatient files, but if a CPT code was not found in the outpatient files, then the calculated total paid value from the carrier file was used. For pharmaceutical costs, we used market history to identify wholesale acquisition cost prices from ProspectoRx. For ED costs, we obtained costs from the literature [12, 13] and used the mean of both UTI-related and all-cause visits. All costs were adjusted to 2020 Q1 US dollars ($). Additionally, we constructed winsorized cost outcome variables to minimize the effect of abnormally high utilization patients on our estimates. Categorical costs, such as all-cause drug costs in the follow-up window, were set to the 98th percentile values. In other words, for patients with costs below the 98th percentile, winsorized costs were set to actual cost; those with costs above the 98th percentile had winsorized cost set to 98th percentile cost. The goal of this procedure was to reduce the influence of outlier observations. All-cause services included the cost of all services provided to patients with uUTI in the 6 months after the index uUTI event, while UTI-related costs included the costs of services provided where a uUTI diagnosis code was included in the service event in the encounter file or the service was a uUTI-related lab/culture.
Statistical analyses
For the primary outcome measure, to evaluate how infection with not-susceptible isolates affected healthcare utilization and costs, propensity score matching was used to align the populations in terms of the following covariates: age; White race; private insurance; inpatient/outpatient clinic/outpatient other/ED visits within 180 days pre-index; and all-cause drug orders within 180 days pre-index. Patients with uUTIs with isolates not-susceptible to ≥ 1 antibiotic were matched with those with susceptible urine isolates using 1:1 nearest neighbor propensity score matching to minimize the difference across the pairs. As cost data are typically right-skewed, we used a generalized linear model (GLM) with a log link and gamma family to account for this skewness in the dependent variable; we split the model into an analysis of the effect of resistance on the extensive margin and intensive margin, individually. Patients retained for analysis had a non-zero predicted probability of being in the case and control group and were retained for analysis only if there were patients in the mirror group with similar propensity scores. The propensity score was calculated using a logistic regression.
Secondary analyses used the same propensity score matching approach as for the primary analysis. The difference in cost outcomes of interest was measured using the same GLM with a log link and gamma family to adjust for patient characteristics. When calculating how antibiotic resistance affected healthcare utilization, a negative binomial model was used to analyze the effect of antibiotic resistance on the following discrete outcome variables: the probability of having at least one all-cause/UTI-related inpatient visit; the probability of having at least one all-cause/UTI-related outpatient visit; or the probability of having at least one all-cause/UTI-related ED visit.
Logistic regression specification was used to estimate the likelihood of progressing to cUTI. Patients were identified as having progressed to cUTI if they had a new symptom for fever, nausea, vomiting, or received a new administration of IV antibiotics between 3 and 28 days post-index uUTI event. Following this, 6-month cost and utilization numbers for patients with uUTI who progressed to cUTI within 28 days of the index uUTI diagnosis were compared with those who did not progress. The same logistic regression models were used to estimate the prediction of the treatment failure outcomes given a patient’s antibiotic resistance status. Logistic regression models were estimated only on the unmatched baseline population of patients who could achieve a given outcome (e.g., the population used in the analysis of the probability of failure of a first-line therapy was restricted to patients who received a first-line therapy as their index antibiotic).
Sensitivity analyses were performed to test the robustness of the primary analysis. In sensitivity analysis #1, patients were required to have both a diagnosis code and positive urine cultures for uUTI inclusion, applying the same 6-month follow-up window as the baseline analysis, to make the population smaller and increase confidence that only patients with uUTI were being analyzed. Sensitivity analysis #2 defined a 30-day post-index uUTI follow-up window that was compared with the 6-month window used to measure all-cause/UTI-related cost outcomes, where previously this was used to measure utilization costs only. Sensitivity analysis #3 used a 12-month post-index uUTI follow-up window to measure utilization and cost of the primary and secondary outcomes, respectively, in order to better capture the long-term effects of infections attributed to antibiotic resistance. In sensitivity analysis #4, the exclusion criteria were simplified and reduced per a previous study [14] to have a less restrictive uUTI study population. Sensitivity analysis #5 was designed to measure the difference in UTI-related and total healthcare resource utilization and costs to compare patients with recurrent uUTIs to patients with one episode of uUTI. In sensitivity analysis #6, patients with concurrent non-uUTI conditions who could be treated with an antibiotic were excluded to ensure infectious conditions with major economic burden were not overinflating costs. Finally, sensitivity analysis #7 was designed to investigate whether the majority of the inappropriate prescribing practices were related to fluoroquinolones (FQs), as their use is discouraged by IDSA guidelines [3]. We repeated the primary analysis but limited the sample to patients receiving FQs as first-line therapy.
