Database and study population

The Clinical Practice Research Datalink (CPRD) database is a nationally representative database of primary care consultations in the United Kingdom [15]. Data in CPRD are collected anonymously from practice management systems of 674 practices and include demographic information, medical tests, diagnoses, and prescriptions. Diagnoses are entered directly by clinicians using Read codes, the main medical coding terminology in UK primary care [16]. A subset of consenting English patients and practices (75% of English practices, 58% of all practices) are further linked to data on hospital admissions and visits to the Emergency Department from the Hospital Episode Statistics (HES) and census data from the Office for National Statistics (ONS).

We included all patients in the CPRD-HES-ONS linked data aged 65 years or more between April 1, 2007, and March 31, 2015. Data were required to fulfil basic quality standards [15], and patients entered the cohort at the latest of the practice’s up-to-standard date, 1 year of continuous registration with the practice, their 65th birthday, or April 1, 2007. Patients left the cohort either on their date of death or 60 days before the earliest of the practice’s last collection date, their transfer-out date, or March 31, 2015. All included patients had a minimum of 60 days follow-up, with the exception of patients who died, because excluding these individuals would bias our results.

Ethical approval

This study was conducted as part of the Preserving Antibiotics through Safe Stewardship (PASS) project [17]. Access to CPRD data within PASS was approved by the Medicines and Healthcare products Regulatory Agency (MHRA UK) Independent Scientific Advisory Committee (ISAC-Nr.: 17 048), under Section 251 (UK National Health Service [NHS] Social Care Act 2006). Individual patient consent was not required for this observational study of anonymized data.

Definition of UTI episodes

The study population comprised patients who consulted for a new episode of lower UTI that originated in the community. Primary care consultation for community-onset lower UTI was identified from the primary care record based on Read codes using previously published codelists [13] (S1 Table) and supplemented with data from the linked hospital record based on International Classification of Diseases 10th revision (ICD-10) codes to exclude cases that originated in hospital. A major challenge in the analysis of routine data is distinguishing between new and ongoing episodes of infection and differentiating between community-onset and hospital-onset infections. For example, a patient could consult primary care for urinary symptoms twice in a 3-month period and depending on the interval between events, this might be classified as 1 or 2 episodes of UTI. Similarly, a patient may be diagnosed with UTI in hospital but consult primary care for urinary symptoms 2 weeks later following discharge. Here, the primary care event is likely to represent continuation of a UTI that originated in hospital.

For these reasons, we applied a strict definition of UTI to restrict our analysis to community-onset cases of lower UTI and to differentiate between new and ongoing UTI episodes. For each patient, the first episode of infection was defined as the date of the earliest observed UTI code in primary and/or secondary care (Fig 1). Like Gharbi and colleagues, we considered any UTI codes recorded in the following 60 days to be part of the same episode. However, whereas Gharbi and colleagues used a fixed time period to define a UTI episode, we applied rules to differentiate between new and ongoing episodes of UTI, based on patterns of consultation for UTI recorded during the 60-day follow-up period (Fig 1A and Fig 1B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Classification of UTI episodes for 2 scenarios for a patient with 3 records of UTI, which are identical except for the timing of the second UTI code.

In both panels, the first UTI code marks the start of a new UTI episode (first episode). The second UTI code occurs within 60 days and is therefore considered to be part of the first episode. The third UTI code occurs more than 60 days after the start of the first episode and is classified as (A) a new episode (because the last evidence of UTI was recorded more than 60 days earlier); (B) an ongoing episode that is excluded from the analysis (because the last evidence of UTI, i.e., second UTI code, was recorded less than 60 days before and may therefore represent an ongoing episode of infection). UTI, urinary tract infection.

https://doi.org/10.1371/journal.pmed.1003336.g001

We further excluded episodes in which the patient had any of the following outcomes recorded on the same day: hospital admission, accident and emergency (A&E) attendance, referral to specialist care, and death. Episodes were also excluded if the linked HES record showed that the patient was in hospital on the date that the episode was recorded in primary care.

Exposure, outcomes, and covariates

We compared patients who were immediately treated with antibiotics defined as prescription of systemic antibiotics on the same day as the episode start date to patients who were not treated with antibiotics on the same day. In contrast to Gharbi and colleagues, we considered patients who were not prescribed antibiotics and those with a delayed prescription (i.e., antibiotics prescribed in the 7 days after—but not including—the day of initial consultation) as a single group, because delayed antibiotic prescribing is not well recorded in EHRs [18].

The primary outcome was BSI within 60 days of the episode start date recorded in the primary or secondary care record. Although the terms sepsis and BSI are not interchangeable, ICD-10 diagnostic codes usually record “sepsis” [19] rather than BSI, even in cases with a positive microbial culture of blood [20]. We have therefore interpreted an ICD-10 code for sepsis as evidence of BSI and used the term BSI throughout. Secondary outcomes were all-cause mortality within 60 days, admission to hospital for reasons unrelated to UTI or BSI within 60 days, and underlying cause of BSI. A 60-day follow-up for these outcomes was selected for purposes of comparison with the literature, notably Gharbi and colleagues. BSI was identified in primary care using Read codes and in secondary care using ICD-10 codes (which represent the primary and secondary reasons for admission) using published codelists [13]. ICD-10 codes for sepsis were further classified as urosepsis, sepsis of other infectious origin, and unspecified sepsis (S1 Text).

Explanatory variables included demographic characteristics: age at episode start, gender, quintile of socioeconomic status (Index of Multiple Deprivation [IMD] 2015), and practice region (South of England, London, East of England and Midlands, North of England and Yorkshire). We also evaluated risk factors for infection and healthcare utilization including Charlson Comorbidity Index (CCI), smoking status (nonsmoker, ex-smoker, current smoker), whether the index consultation was performed as a home visit, recent hospitalizations (discharge in prior 7 and 30 days, number of admissions in prior year, total number of days spent in hospital in prior year), recent A&E attendances (attendance in prior 30 days, number of attendances in prior year), and prescription of systemic antibiotics in primary care in prior 30 days. History of recurrent UTI was defined as an explicit code for recurrent UTI, a prescription of nitrofurantoin or trimethoprim for 28 days or more (prophylactic treatment), or 2 or more consultations for UTI within a year of episode start [13]. CCI and smoking status were calculated using all medical history in primary care before the episode start date. Patients without a smoking code were considered nonsmokers. Patients whose latest record indicated a nonsmoker but who had a previous record of smoking were classified as ex-smokers.

Statistical analysis

We undertook a univariable analysis comparing patients with and without immediate antibiotic treatment for each included variable. Continuous variables were summarized using means median and interquartile range (IQR), and categorical variables using absolute numbers and proportions. Wilcoxon rank-sum tests (continuous) and χ² tests (categorical) were used to assess the difference between exposure groups. We tabulated diagnostic information relating to the underlying cause of BSI for each treatment group.

Crude associations (odds ratios [ORs]) between each included variable and BSI were estimated using generalized estimating equations (GEEs) with a logit link and an exchangeable correlation structure to account for multiple UTI episodes per patient. All count variables (CCI, number of admissions, number of days spent in hospital, and number of A&E attendances) were transformed using the square root before adding them to the GEEs. Huber-White sandwich estimators were used to calculate 95% confidence intervals (95% CI). A final multivariable adjusted model was fitted, including all predictors with a p-value < 0.2 in the univariable analysis. Based on reviewer comments, interactions between prescribing and age or gender were also considered. The number needed to be exposed (i.e., not treated with antibiotics) to harm (NNEH) was calculated from the final model using average risk difference to adjust for covariate imbalance [21]. The analysis was refitted in 200 bootstrapped samples to estimate 95% CIs for the NNEH.

The same approach was used for secondary outcomes. Sensitivity analysis was undertaken restricting the follow-up/wash-out periods to 30 days and only including the first UTI episode per patient. We also tested the sensitivity to residual confounding by performing propensity score analysis. A patient’s prior likelihood to receive treatment was estimated using multivariable logistic regression (parametric) or generalized boosted regression (nonparametric), and 4 different adjusted results were obtained using each set of propensity scores with either matching or inverse probability weighting.

The analysis presented here was outlined prospectively in the protocol submitted to the MHRA Independent Scientific Advisory Committee for ethical approval (S1 Protocol). The definitions and methods were chosen to replicate the analysis performed by Gharbi and colleagues [13] as closely as possible and to address all concerns raised by researchers regarding the validity of those findings [14]. Analysis was performed using the statistical software R version 3.6.1 for Windows [22]. Generalized estimating equations were fitted using geepack (version 1.2–1), and propensity score analysis was performed using MatchIt (version 3.0.2) and twang (version 1.5). This study is reported as per the REporting of studies Conducted using Observational Routinely collected Data (RECORD) guideline (S1 RECORD Checklist). Code for all analyses can be found at https://github.com/prockenschaub/CPRD_UTI_sepsis_elderly.

Comparison with existing literature

Previous studies of alternatives to antibiotics or delayed prescribing for community-onset UTI have usually focused on women aged 18–70 years. A systematic review of trials in young, nonpregnant women reported that antibiotic treatment was associated with more rapid resolution of urinary symptoms and microbiological cure based on urine culture, compared with placebo [24], but not with reduced incidence of pyelonephritis. Delayed prescribing has also been safely used in low-risk women with uncomplicated UTI [25], provided there is adequate safety-netting and self-care advice [2]. For example, in a trial comparing treatments for uncomplicated UTI in women aged 18–65 years [26], women receiving ibuprofen had a higher burden of symptoms but considerably less antibiotic exposure compared with women treated with fosfomycin (incident risk reduction 66.5%, 95% CI 58.8%–74.4%; p < 0.001), and two-thirds of patients in the ibuprofen group recovered without antibiotics.

Although the efficacy and safety of delayed prescribing for respiratory tract infections in primary care is well-established [27], implementing similar approaches for UTI is controversial because of concerns around prolongation of symptoms and the potential risk of antimicrobial resistance and complicated UTI resulting from inadequate antibiotic therapy. These issues are particularly relevant in elderly patients who have the highest incidence of community-onset UTI, but also the highest incidence of E.coli BSI [3], which may be a consequence of suboptimal antibiotic treatment in primary care. However, it is unclear whether use of delayed prescribing for suspected UTI would be acceptable to GPs or to patients in this age group.

With the exception of Gharbi and colleagues, few studies have evaluated the use of delayed prescribing or alternatives to antibiotics in older adults. These patients arguably have the most to gain from prudent antibiotic prescribing because of their increased risk of adverse outcomes related to antibiotic use [28] and high prevalence of asymptomatic bacteriuria [6]. The major barrier to delaying or withholding antibiotics in these individuals is the risk of UTI-related complications, as reported by Gharbi and colleagues [13]. Our analysis, and concerns raised by other research groups [14], call these findings into question. Recognizing the limitations of analyses based on routine data, we find no evidence of an association between delaying or withholding antibiotics and bloodstream infection but some evidence of increased mortality. The discrepancy between our analysis and that conducted by Gharbi and colleagues is likely to relate to the different approaches used to define community-onset UTI and the limitations of coding in EHRs. In our adjusted analysis, the biggest reduction in effect sizes related to inclusion of information on home visits, possibly because prescribing outside the practice is not recorded electronically, and greater comparability of exposure groups, due to exclusion of cases that did not meet criteria for “community-onset.”

Strengths and limitations of this study

A major strength of our analysis is the use of a large and nationally representative primary care database (>850,000 patients) linked to hospital admissions. This means our estimates can be generalized to the UK population [15]. Linkage of the primary care dataset to HES allowed us to apply stringent criteria to identify community-onset UTI cases by differentiating new from ongoing UTI episodes and excluding cases that originated in hospital. Sensitivity analyses also support our main conclusions and highlight the limitations of diagnostic coding for BSI.

Limitations relate to the fact that EHRs are designed for clinical care not research. Observational studies using CPRD are at risk of confounding by indication if there are systematic differences (such as the severity of symptoms) between patients who receive a prescription and those who do not. This is particularly challenging when the exposure of interest is unevenly distributed across the study population as seen in this study (87% of patients received an immediate antibiotic versus 13% who did not). Estimates from our propensity score analysis were congruent with our main findings, but we acknowledge that residual confounding is likely. Remaining differences in patient characteristics between groups suggest that we were unable to fully account for confounders. We conclude that residual confounding persists, likely due to factors that influence GPs prescribing decisions but are not well recorded in EHRs, such as severity of clinical presentation; patient’s prior medical history; knowledge of patient preferences, for example, in relation to end of life care; and the patient’s social circumstances.

Read codes were used to identify patients with suspected UTI, because microbiological culture of urine is usually only performed for patients with recurrent UTI or when the clinician suspects that the patients may have a drug-resistant infection. Consequently, it is likely that our cohort included patients with asymptomatic bacteriuria and/or other types of infections. Up to 40% of prescriptions for nitrofurantoin are not linked to a Read code [23], which suggests that we may have failed to identify some patients who were treated immediately with antibiotics. This also highlights challenges associated with using Read codes to infer the date of infection onset. Similarly, we made the assumption that patients commenced their antibiotic treatment on the date the prescription was issued.

Because delayed prescribing is not well recorded in primary care, it is feasible that some individuals recorded as receiving antibiotics immediately actually received a delayed prescription. Depending on the extent of misclassification between treatment groups, this may have led us to underestimate the reported association between antibiotic treatment and adverse outcomes. However, whereas delayed prescribing is commonly used for respiratory tract infections, evidence to support this approach for UTI is limited. For these reasons, it seems likely that most who were prescribed antibiotics on the date of their consultation started treatment on the same date. Conversely, if there were substantial differences in outcome between delayed prescribing and no prescribing, treating them as 1 group may overestimate the reported association between delayed prescribing and mortality, although the results reported by Gharbi and colleagues [13] suggests that this is not the case. Finally, we used the CCI as a composite measure of comorbidity. This had the advantage of making our analysis comparable with Gharbi and colleagues [13], but it does not take account of the fact that specific comorbidities, such as those affecting the urogenital tract, may impact an individual’s risk of BSI more than others (and therefore influence GPs’ prescribing decisions).

Cases of sepsis were identified from ICD-10 codes or Read codes, and we found that almost one-third of sepsis diagnoses were only recorded in primary care. It is difficult to disentangle the reasons for this because almost all cases of sepsis are managed in hospital. Patients may have received treatment for sepsis abroad or in a non-NHS setting, or information from the discharge letter may have been used to infer the diagnosis of sepsis. Linkage of microbiological data to HES/CPRD would enable more accurate estimation of the proportion of BSI cases that could be attributed to a urinary source and resolve questions around the proportion of cases with a “sepsis” diagnostic code who have microbiological evidence of BSI.

Clinical, policy, and research implications

This population-based study highlights uncertainty around the risks and benefits of antibiotic treatment for suspected UTI in patients aged ≥65 years. The increased risk of adverse outcomes in this age-group may make GPs more likely to prescribe antibiotics, increasing the likelihood that these patients will be exposed to antibiotics unnecessarily. For researchers, our findings highlight methodological challenges associated with defining the onset of infection and addressing confounding when analyzing EHRs and the need for linkage of microbiological datasets to HES/CPRD. There is also a need for qualitative research to understand patients’ and GPs’ views on the acceptability of delayed prescribing for UTI in this age group to inform the design of future studies.