Safety of reducing antibiotic prescribing in primary care: a mixed-methods study

Ethics statement

The proposal for the qualitative research study was reviewed and approved by the London – Hampstead NHS Research Ethics Committee (reference 18/LO/1874).

Study design

Semistructured interviews were conducted with patients and primary care prescribers, including GPs, nurses and pharmacists, in two English regions (one urban metropolitan area and one shire town with a high demand for primary care services). Metropolitan practices were invited to the study by the local Clinical Research Network that generated the expression of interest. A shire town practice was recruited through informal Clinical Research Network contact, which also helped to liase with potential respondents.

Recruitment of participants and data collection

The research invitations to participate generated expressions of interest from the general practices that agreed to purposively select patients who visited a primary care professional for infection in the last 6 months. Patient lists were approved by a GP acting as research gatekeeper in each practice and, initially, 927 patients were sent invitations to the study via the Docmail^® postal system (Radstock, UK). The invitations contained a letter from the practice, inviting patients to participate, and an information sheet. Patients who agreed to take part either returned reply slips or contacted the researcher using the contact details provided. The researcher then communicated via e-mail or text message to establish the contact, followed by sending the consent form and confirmation of the interview meeting.

The sample size was determined using the pragmatic concept of ‘information power’,66 which proposes that the size of a sample with sufficient information power depends on the aim of the study, sample specificity, use of established theory, quality of dialogue and analysis strategy. Although our aim was broad, specificity was biased towards one group (almost half of the interviewees were older female patients consulted for UTI); therefore, we followed a theoretical model to explain the findings and the quality of the interviews was relatively high. As we aimed for a cross-case analysis, we decided to continue recruitment until the sample size reached 31 eligible patients.

Analysis

The interviews were digitally recorded, transcribed verbatim by a professional transcriber, imported to an NVivo 12 project (QSR International, Warrington, UK) and coded through an iterative six-phase process described in thematic analysis. 67 Data analysis occurred iteratively and involved familiarisation, coding, theme searching, theme reviewing, theme defining and naming, and producing the report. Repeated patterns in the data formed the basis for the codes, which were identified by the lead analyst (OB), and one single code for every different concept/idea was generated. To ensure that codes were applied consistently, a co-author (CB) independently coded a random sample of four interview transcripts. Coding was refined after discussion. Data identified by the same code were collated and all different codes were sorted into potential subthemes and themes using NVivo options of tree building. The potential themes were reassessed and reorganised to reflect major narratives and themes in the coded data. Finally, the analysts refined and named the five main themes and subthemes. Participants’ feedback on the transcripts or the summarised final findings was not sought because this was not feasible. However, the main themes were discussed at a patient and public involvement (PPI) meeting, as noted in Patient and public involvement.

Interviews

The practitioner study investigated how primary care prescribers perceive risk and safety concerns associated with reduced antibiotic prescribing. Items for the interview were guided by the theoretical domains framework (TDF), which uses theories of behaviour change to understand factors influencing health-care practice. 68,69 The TDF comprises 14 domains, covering the main theoretical determinants of behaviour, and the interview was designed to reflect these domains (Box 2). 68 The interview was piloted with three GPs to ensure that the questions were appropriate, readily understood and covered relevant prescribing behaviours. All interviews were conducted by Olga Boiko to ensure consistent quality. All interviews except one telephone interview were conducted face to face on general practice (n = 26) and university (n = 4) premises in the period January–July 2019. The interviews lasted between 24 and 46 minutes. The participants were offered £60 to acknowledge their contribution.

Recruitment of participants

Contact details of primary care practitioners were obtained through the help of the local research facilitators and practice managers. Potential participants were then approached either directly by e-mail using the study information pack or indirectly via the practice manager or lead GP. The information pack included the invitation letter and study information sheet. A reminder was sent out 2 weeks after the initial approach to those who had not responded. A purposive sampling approach was followed. Forty-nine primary care providers from 10 general practices were invited and 30 agreed to take part. This number of participants was deemed sufficient, using the pragmatic concept of ‘information power’. 66 The uptake varied between practices (in five practices only a single primary care practitioner was interviewed). Interviews took place between January 2019 and July 2019.

Analysis

The interviews were digitally recorded, transcribed by a professional transcriber, imported to an NVivo 12 project and coded through an iterative six-phase process, as outlined above. 67 Repeated patterns in the data formed the basis for the codes, identified by the first author (OB), and one single code for every different concept/idea was generated. To ensure that codes were applied consistently, a co-author (CB) independently coded a random sample of four interview transcripts. Coding was refined after discussion. Data identified by the same code were collated and all different codes were sorted into potential subthemes and themes using NVivo options of tree building. Next, the potential themes were reassessed and reorganised to reflect major narratives and themes in the coded data. Finally, the themes and subthemes were refined and named (by OB, CB and MCG). The themes and subthemes were then mapped to the relevant domains of the TDF to assess the relative importance and salience of individual domains.

Ethics statement

The protocol for the epidemiological study was approved by the CPRD Independent Scientific Advisory Committee (protocols 18-041R and 19_110R). The CPRD holds overarching Multicentre Research Ethics Committee approval for the database and conduct of studies using fully anonymised data.

Data source

We carried out population-based cohort studies in the UK CPRD GOLD database, employing data for 2002–17. CPRD GOLD draws on general practices that use the Vision general practice software system. The Vision system has suffered from a declining market share in recent years. Consequently, CPRD has established the CPRD Aurum database, which draws data from general practices that use the EMIS general practice system.

The CPRD GOLD database collects data from the four countries of the UK, with about 30% of contributing practices located in England at the time of this study. The CPRD GOLD database has been well described70 and the high quality of the data collected has been documented in many studies. 71 CPRD records include details of consultations by general practice staff, as well as coded records of referrals to hospital or discharge letters from hospitals. The CPRD GOLD is a ‘live’ database, which is updated monthly. The October 2019 database release included data on 17.6 million patients, of whom 2.6 million were currently active. There were 320 general practices currently contributing to CPRD GOLD: 30% in England, 3% in Northern Ireland, 37% in Scotland and 30% in Wales. Data linkage in CPRD GOLD is restricted to volunteer general practices in England only. The CPRD Aurum database was more recently established and, at the time of this study (June 2019 release), was restricted to data collected from general practices in England. 72 The CPRD Aurum database included data on 883 general practices, from which patients were sampled, with 23.1 million patients, including 2.5 million currently active patients. This study was designed using the CPRD GOLD database. However, we conducted a substudy to compare antibiotic prescribing metrics between CPRD GOLD and CPRD Aurum.

The research aimed to evaluate safety outcomes of reduced antibiotic prescribing, including conditions such as sepsis, PTA and infective endocarditis. These are infrequent or rare events. Consequently, it was necessary to evaluate outcomes over a long period of time in the whole CPRD database to obtain sufficiently precise estimates of incidence rates. The period from 2002 to 2017 was selected for study. However, our licence agreement with CPRD places limits on the number of records that can be extracted for analysis. We were able to extract full CPRD data for the numerator, but for the denominator we were restricted to data included in the CPRD GOLD denominator file, which comprised age (year of birth), gender and study year. To address this, we employed sample data for the study denominator, as outlined below.

The research was further complicated by possible changes over time in the definition and recognition of outcomes. For example, definitions of sepsis have changed over time, and there have been substantial changes in professional and public awareness of sepsis as a complication of infection. There have also been changes in approach to antibiotic prescribing and antimicrobial stewardship during the period of study. This required analytical approaches that accounted for changes over time in key exposures or the incidence of outcomes.

As noted above, the CPRD GOLD is a ‘live’ database that is updated each month. The ‘last collection date’ for each general practice is updated for each release, as are the dates of death and the end of registration for each participant. In addition, the ‘up-to-standard’ date at which the general practice is judged to have been contributing research-quality data may be updated, even for historical data, based on an algorithm employed by CPRD. Patients have the possibility of ‘opting out’ of CPRD and the small number of opting-out patients may change over time. The present research was conducted over several years in the form of a series of related studies. The research is, therefore, based on data from several different releases of CPRD GOLD. In addition, we made ongoing revisions and updates to our medical code lists to enable updating consistent with our evolving understanding of conditions relevant for study. Consequently, the number of outcome events and person-years at risk may vary slightly when different analyses are compared, although relevant findings are expected to be consistent across different releases.

Antibiotic prescribing study

We evaluated rates for antibiotic prescription and infection consultations in the CPRD GOLD database between 2002 and 2017.

Comparison of antibiotic prescribing in Clinical Practice Research Datalink GOLD and Clinical Practice Research Datalink Aurum

To evaluate the transferability of our findings with respect to antibiotic prescribing between general practices using the Vision and EMIS practice systems, we conducted a substudy to compare estimates using sample data from the CPRD GOLD and CPRD Aurum databases. We evaluated antibiotic prescribing for the year 2017, as this was the most recent complete year for our study. 39

General practice-level analysis of serious bacterial infections

Incident cases of serious bacterial infection were evaluated in the January 2019 release of CPRD for the years 2002–17, with the CPRD denominator providing the person-time at risk. The mean duration of follow-up was 6.9 years. Serious bacterial infections were selected for study from review of the International Statistical Classification of Diseases and Related Health Problems, Tenth Revision (ICD-10),77 from the Read code classification78 and through discussion with the research team. The final list of conditions is summarised in Table 2 and includes bacterial infections of the central nervous system (CNS), bacterial infections of the cardiovascular system (CVS), kidney infections, lung abscess and empyema, mastoiditis, osteomyelitis, PTA, resistant infections and Clostridium difficile (C. difficile), sepsis and septic arthritis. Further details of case definitions are provided in Report Supplementary Material 1 and 2. Incident events were the first record for each type of serious bacterial infection in a patient > 12 months after the start of the patient record. However, a single patient might have first episodes of more than one type of bacterial infection. Possible recurrent events in the same patient were not evaluated further because, in electronic health records, it may not be possible to distinguish new occurrences from reference to ongoing or previous problems.

Serious bacterial infections were analysed as the outcome (Equation 2):

The antibiotic prescribing level for each general practice was included as a predictor using the general practice-specific estimates from Equation 1. These estimates initially had a mean of zero and a standard deviation of 0.19, which is consistent with an adjusted relative rate of antibiotic prescribing of 1.21 for a general practice prescribing 1 standard deviation above the mean. Estimates were, therefore, standardised to give the change in serious bacterial infection for a 20% relative increase in antibiotic prescribing rate at a practice, as this represents a change of approximately 1 standard deviation. A 20% change generally represents a substantial change in antibiotic prescribing. We also estimated the change in serious bacterial infection for a 20% relative increase in the proportion of antibiotic prescriptions, with specific medical codes recorded at a general practice. Models were adjusted for age group, gender, region, deprivation fifth and calendar year (including quadratic and cubic terms for calendar year), with the log of person-time as offset (see Equation 2). The results were visualised using forest plots.

Decision tree

To evaluate the probability of a serious bacterial infection following an infection consultation in primary care we constructed a decision tree (Figure 3). 79 An individual developing an infection may decide to consult their general practice or not. If they consult, they then may be prescribed antibiotics and subsequently they may develop a serious bacterial infection. We used estimates from CPRD data analysis to populate the decision tree with empirical estimates for probabilities, as outlined in Table 3. We used Bayes’ theorem to estimate the probability of a serious bacterial infection following an infection consultation if antibiotics were prescribed or if antibiotics were not prescribed. We estimated the number needed to treat (NNT) (i.e. the number of antibiotic prescriptions required to prevent one serious bacterial infection) as the reciprocal of the difference in probability of sepsis with and without antibiotics. We obtained central estimates and 95% uncertainty intervals (UIs) from 10,000 random draws from the beta distribution. 80 All estimates were stratified by gender and 10-year age group. For the population aged ≥ 55 years, we also stratified by frailty category. In addition, we evaluated subgroups of common infections, including RTIs, skin infections and UTIs.

FIGURE 3.

Decision tree showing the probability of a patient consulting for an infection, being prescribed an antibiotic at that consultation and developing serious bacterial infection (see Table 3 for explanation of abbreviations). AB, antibiotic; SBI, serious bacterial infection.

Data sources for decision model

Study population and data sources

The study employed the UK CPRD GOLD database. The CPRD GOLD is a primary care database of anonymised electronic health records for general practices in the UK. The high quality of CPRD GOLD data is well established. 71 CPRD GOLD has a coverage of some 11.3 million patients, including approximately 7% of the UK population, of which it is broadly representative in terms of age and sex. 70 Consenting practices in England participate in a data linkage scheme. 82 Approximately 74% of all CPRD GOLD practices in England are eligible for linkage. Linkages are available for the HES and mortality registration data from the Office for National Statistics (ONS). HES admitted patient care data include admission and discharge dates and diagnostic data coded using the ICD-10. Mortality registration data include information on the date and causes of death coded using ICD-10. ONS identifies one underlying cause of death and secondary causes of death, including up to 15 additional causes of death. Linked area-based measures of deprivation include the Index of Multiple Deprivation (IMD) and are based on a weighted profile of indicators. 83 We employed deprivation for the general practice postcode for this study because of the low proportion of missing values. The protocol was approved by the CPRD Independent Scientific Advisory Committee (protocol 18-041R).

Main measures

We included patient records between 1 January 2002 and 31 December 2017. The start of the patient record was the later of the patient registration date or the date that the general practice joined CPRD. The end of the patient record was the earliest of the last data collection date, the end of registration or the date of death. We evaluated the first records of sepsis > 12 months after the start of registration in primary care electronic health records as a primary diagnosis in HES or sepsis as any mentioned cause of death in mortality records. In UK primary care records, diagnoses recorded at consultations or referrals to or from hospitals were coded, at the time of this study, using Read codes. We identified sepsis records using a list of 77 eligible Read codes. Incident episodes of sepsis in CPRD were recorded using 55 Read codes, with four codes accounting for 92% of events, including ‘sepsis’ (two codes) (64%), ‘septicaemia’ (18%) and ‘urosepsis’ (10%). In HES and death registry records, sepsis diagnoses and sepsis deaths were defined using 23 ICD-10 codes for sepsis. In HES records, we evaluated the primary diagnosis, which accounts for the majority of the length of stay of the episode, with other diagnoses being referred to as comorbidities. 84 Incident diagnoses of sepsis in HES were coded with 20 ICD-10 codes, with three codes accounting for 89% of events, including ‘sepsis, unspecified’ (72%), ‘sepsis due to other Gram-negative organisms’ (13%) and ‘sepsis due to Staphylococcus aureus’ (5%). In mortality data, we included all mentioned causes of death because sepsis may be part of a sequence of morbid events and not always be an underlying cause of death. 24 ‘Sepsis, unspecified’ accounted for 93% of causes of death among those in the ONS death registry with sepsis as any mentioned cause of death.

Analysis

Incident sepsis events were identified for each data source. We calculated person-time at risk from the start to the end of the patient record. Person-time was grouped by gender and age group from 0 to 4 years, 5 to 9 years and 10 to 14 years, and then 10-year age groups up to ≥ 85 years. Incidence and mortality rates were age standardised using the European Standard Population for reference. We searched for concurrent events across data sources using a 30-day time window. We calculated age-specific incidence rates using primary care electronic health records and then adding HES records, mortality records or both. We fitted a logistic regression model to evaluate associations of gender, age group, fifth of deprivation and period of diagnosis with concurrent sepsis recording. All data were analysed in R.

Sensitivity analyses

To consider recurrent sepsis events, we conducted a sensitivity analysis using CPRD and HES where incident events were first sepsis records during each calendar year during the study period. We also evaluated the effect of extending the time window for concurrent events from 30 to 90 days.

Design of study

The study used a qualitative design involving both semistructured and ‘think-aloud’ interviews with six GPs. The study was designed and conducted by LM in consultation with the study team. Face-to-face interviews were conducted online using Zoom (Zoom Video Communications, Inc., San Jose, CA, USA). Interviews lasted approximately 40 minutes. All interviews were recorded using Zoom and fully transcribed. The evaluation was constrained by the circumstances of the research during the COVID-19 pandemic and it was possible to conduct a preliminary study only.

Participants

Participants for early user testing were six GPs from practices across London, Southampton and Oxfordshire. Participants included GPs who were also academic researchers and GPs who worked in a practice setting only. Participants were recruited initially from members of the study team who were also practising GPs. In addition, two GPs who were colleagues of study members also agreed to take part following an invitation from the study team. This part of the research was conducted during January–July 2020 and its scope was limited by the circumstances of the research.

Procedure

General practitioners were first shown the web pages using the screen share function on Zoom. Think-aloud interviews (Box 3) were conducted at this point to study the reactions to the various features of the web pages. GPs were encouraged to discuss their views, thoughts and perceptions of each section as it was demonstrated by the interviewer. This technique allowed GPs to explore the various drop-down menu options and openly discuss the tool as they wished, but also ensured that opinions were obtained for all sections of the pages. Following this, a semistructured interview (Box 4) was conducted, drawing on our previous experience of developing electronic interventions. 31,85 The interview was designed to identify factors likely to influence successful implementation of the tool and discover likely responses to the proposed sections to further inform development and aid refinement of the web pages. GPs were asked questions regarding their views, expectations, acceptability and feasibility of pages. The semistructured interviews were used to explore and discover issues that may be related to the web page content and usage.

Analysis

Inductive thematic analysis was conducted on all transcripts to determine likely responses to the web pages and identify factors involved in the decision to use the tool. Analysis began after the first interview had been conducted and continued throughout data collection for all interviews conducted. Interviews were read in detail and re-read, and then following this immersion in the transcripts commonly occurring patterns and prominent themes were identified in the data and labelled with codes. Each code label referred to the operationalisation of the theme content. A coding manual was developed containing the label, a definition of each theme, positive examples from the interview transcripts and possible exclusions for each code. The coding manual was refined as more data became available and transcripts were re-read. The continuing process involved themes being linked, grouped, moved, re-labelled, added and removed to produce a set of themes and coding manual that adequately fit and thoroughly explained the data.

Beliefs about antibiotics and antimicrobial resistance

Antibiotics emerged as trusted medicines that had widespread use. The descriptors used by the interviewed patients ranged from ‘magic answer’ to ‘sledgehammer’ treatment. Those who referred to antibiotics as a ‘magic’ pill were often older females with recurrent UTIs and with expectations for apparently appropriate prescribing. On a whole, the most common belief among the interviewees was that antibiotics should be prescribed and taken when necessary. The patients’ concerns were rather about ‘finding the right one for what infection you have at the time and making sure that you’re going to be safe’ (interview 13, female, chest infection). There was also recognition of the need for better scrutiny in prescribing, which highlighted the complexity of decision-making concerning safe antibiotic treatment.

The interviewees spoke about different thresholds of illness and about reducing the use of antibiotics. There were accounts of the decreased impact of antibiotics and of antimicrobial resistance by informed patients, especially by younger and more educated patients:

Tell me more about antibiotics. What do you know about antibiotics?

. . . the effectiveness of some of our standard treatments is decreasing rapidly. And I believe that a lot of that is to do with inappropriate usage, both inappropriate prescribing which I have witnessed myself. But then also inappropriate use by patients of antibiotics.

Or in another example:

Do you want to add anything else in terms of your experience with antibiotics?

I find it alarming that we may be getting towards the end of the road with antibiotics . . . They’re finding ways of compensating for the overuse of antibiotics. And clearly there are issues with the pharmaceutical industry as to how much they’re prepared to invest in developing new antibiotics . . . And I don’t want to get political about it but there needs to be some sort of disentanglement of the profit motive, which I understand and the service to the general public.

Expectations of antibiotics

Interviewees’ expectations varied. Approximately half of the interviewees expressed a wish for an antibiotic, but the other half had undifferentiated expectations of help in getting better when attending for the consultation. Those with recurrent UTIs wanted antibiotics based on their previous experiences:

Did you have expectations of specific treatments when you went to see the GP?

While I was waiting for the antibiotics, I also tried stuff myself from the chemist. And also, drunk the cranberry juice, which was no good. I knew that I needed antibiotics. It seems for me if I get cystitis it starts, and it comes on really quickly.

Those with chest infections had thought about the need for antibiotics, but preferred to leave the decision concerning antibiotic treatment to health-care professionals:

When you went to see the doctor, did you expect a particular treatment or prescriptions?

I did expect that if it is something on the lungs, that I would be given antibiotics. But I didn’t push for them or anything. I really went there to, to see what it is. But I wasn’t particularly surprised that they heard the noise on the lungs.

Patients interviewed often referred to antibiotics as something that would ‘shift’ their illness, but also as a symptomatic cure and something ‘to boost the immune system with’. This female patient held a radically different view:

OK, so, you didn’t expect any particular treatments or antibiotics in particular?

I feel bad to say this to the doctor, but I’m quite anti the use of antibiotics. I think I’ve read too many horror stories about overuse of antibiotics. So, I avoid them at all costs. And I definitely wouldn’t have even considered antibiotics unless the doctor had mentioned them.

Approximately half of the interviewees, and those who were not consulting for UTI, had less differentiated expectations:

Did you have any expectations of specific treatment when you came with the symptoms?

I didn’t have any expectation of treatment. And, in actual fact, I remember thinking, ‘Oh do I really need antibiotics?’. But I don’t know, I guess with kind of almost pre-programmed to say, ‘OK, if that’s what you want to give me, that’s what I’ll take’. So, I didn’t really have any expectations. To be honest, I hadn’t even considered that I may have tonsillitis.

In terms of how interviewees perceived prescribers’ expectations, there was a difference in understanding of the prescribers’ mindset. Several interviewees held a view that prescribers’ expectations were contingent on patient expectations and so that they brought up an association with patient pressure: ‘. . . they don’t want to prescribe antibiotics till they see that you are really, really desperate or you need it’ (interview 6, female, tonsillitis). Prescribers’ expectations were in these cases dependent on patients’ wishes as well as on symptoms. Others believed that patient expectations might not influence the professional’s decision, which might depend only on clinical findings. Such views were based on critical reflection:

What do you think doctor thought of, about your expectations?

I don’t think a doctor, any GP in the UK takes into consideration a person’s expectations. They give, they prescribe based on the symptoms that you present. Not necessarily your expectations.

Experience of taking antibiotics

The interviewees had different exposure to antibiotics. Some interviewees were prone to recurrent infections and had been prescribed antibiotics more than once in recent months; whereas others had received an occasional prescription for antibiotics. Most interviewees described days and weeks of experiencing illness before they consulted with a clinician. With the exception of one participant, there had been a sense of welcoming antibiotic treatment; for example, according to a patient with chest infection, ‘antibiotics, without being dramatic, saved my life’ (interview 1). At the same time, many patients were reflexive about the role of antibiotics in coping with the ailments. For example, a patient with tonsillitis questioned the appropriateness of prescribing:

So, it seems like it cleared everything off? Was it efficient?

To be honest, I remember starting to feel like my throat was feeling better within 12 hours of taking the antibiotics. And, at that stage, I started to think, OK, actually maybe I didn’t need antibiotics because this has cleared up very quickly and, in hindsight, kind of thinking about it, thinking, well there’s no way the antibiotics would work that quickly.

The interviewees had a range of various experiences of past antibiotic treatment. Patients with a history of infectious diseases shared positive accounts of antibiotics in general:

So, obviously were you hopeful for the antibiotic treatment to be the best possible course of action when you came with chest infection?

I’ve had chest infections before. They’ve always cleared up and you know I’ve never sort of thought it could lead to anything significantly worse. I’ve always been confident that whatever antibiotic I was given would do the job, you know.

Experience of antimicrobial resistance and side effects

Approximately one-quarter of the patients spoke about their experience of antimicrobial resistance. The accounts were full of frustration and confusion because either the antibiotics had not worked or the first-line treatment had not worked. This was especially true for patients with UTIs, but also for patients with tonsillitis. Several patients experienced up to three to five episodes of UTI per year and found it hard to tolerate its recurrent nature:

Did you mention you had a bad reaction to that as well?

I felt really ill in the morning. I didn’t know why, and I had to drive somewhere. And oh, and really, really had to go to bed. I felt so bad. And then on the Monday as I say, I had a call from the surgery to say they’d discovered that it was resistant to that and they changed it . . . I have heard antibiotics can turn toxic in you if they’re not the right one, can’t they?

Antibiotics were sometimes associated with mild to serious side effects; for example, an anaphylactic allergic reaction was coupled with resistance in the following account:

Did you say you were allergic to one particular antibiotic?

I was allergic, yes.

How did you know that?

I was given penicillin that was in 2000 and I had an anaphylactic shock. I was rushed into hospital and I had to spend a night in the emergency room and given steroids for 5 days because of that. So, then I couldn’t take penicillin, I couldn’t take erythromycin and I became resistant to doxycycline and there were other ones that I started to become resistant to, that didn’t work.

Several interviewees confirmed that the most common side effects were nausea, stomach upset, vomiting and thrush. The interviewees spoke about compensating with probiotics to resolve digestive side effects, but continued with antibiotics. Where discontinued or prescribed different antibiotics, this was in the case of more significant side effects, such as liver derangement or shortness:

So, have you been, having any side effects because of antibiotics?

But when I was prescribed it a second time, within 24 hours I was getting very short of breath . . . it was clear to me that this was having an effect on my breathing. So, I went straight back to the GP and then they put me on, on a different antibiotic.

Experience of consultation

The patients who were interviewed reported predominantly positive experiences of consultation with a prescriber who they had recently seen for infection. With the exception of a very few rushed encounters, when there had not been time for asking/answering questions, most consultations were described as patient centred:

And how did the consultation go?

It was good. She [prescriber] was very thorough. She understood that 3 weeks of suffering was quite a long time, so she was understanding. So, yes, she was thorough, she was very sympathetic, and I felt silly going about something that was really, I thought just a cold. But she said she could tell that I was in discomfort and she listened to my symptoms.

Meanwhile, the patients were concerned with the issue of the appropriateness of antibiotics use and understood the uncertainty and complexity associated with it. They appreciated both diagnostic uncertainty and how diagnostic uncertainty has been resolved in real-time consultations. A patient perceived the complexity of weighing clinical decision-making against the risks to patients’ health:

Who should be making decisions on antibiotics, the doctors or the doctors and patients or maybe patients?

. . . and the doctor would have to be able to take a closer look and to really be able to assess whether he can take the risk of not giving the antibiotics or not, because maybe it’s, it’s just finding a balance of hitting the riskiest bit of this illness first and then dealing with the side effects as you go along.

It was also apparent that the patients who were interviewed reflected on patients’ collective state of mind and role in pressurising a prescriber:

I’m sure many people lie just to get antibiotics.

Interview 19, female, tonsillitis

There were, however, consultations in which shared decision-making took the form of expectation elicitation. Reflexivity emerged in these cases where prescribers directly elicited patient expectations. Two patients admitted being overtly asked about their agreement to use antibiotics:

So, to begin with can I ask you what’s your recent consultation with GP? What was the problem?

. . . she [prescriber] said to me would I mind taking some antibiotics for that. And I said if she felt that they would help then yes, I did. It was a 5-day course . . . Yes, that was the first time actually that I’ve sort of been asked rather than told.

Patient and public involvement

The process of developing subthemes and themes was discussed at a PPI meeting. The research team presented an overview of the interview study, including the objectives and design. The main findings with respect to patient understanding of antibiotics, patient expectations and patient experience of side effects were presented using illustrations from the interview extracts. The meeting discussed selected quotes from patient interviews and the PPI members were asked for their interpretation. The feedback received was included in the final interpretation. The discussion focused on patient pressures for antibiotic prescriptions, trust and communication with GPs. There was discussion of patient education and implied criticism of patient-led consultation if wants and expectations are inconsistent with clinical need. There was also discussion of how dissemination of the findings might contribute to patient education with respect to antibiotic-seeking behaviour.

Main findings and comparison with the literature

Participants perceived antibiotics to be much-needed medicines that should be prescribed when appropriate. Expectations for antibiotic treatment were often conditioned on previous experiences. However, past experiences were not restricted to successful treatment of infections, but also included experience of antibiotic-resistant infections, antibiotic drug side effects and inappropriate prescribing. Patients’ views were also informed by genuine concern about antimicrobial resistance. These lay accounts appeared to reflect the contemporary medical ideas of ‘precision’ or ‘personalised’ medicine, which are represented in the slogan of the ‘right drug for the right patient at the right time’. 104

Consistent with other studies,93 participants’ knowledge about side effects was not associated with their expectations of antibiotics. The concern about and experience of antimicrobial resistance found in our sample contradicts evidence from a review of general public attitudes, which reported low awareness of antimicrobial resistance. 105 We found that around one-quarter of the sample experienced antimicrobial resistance in one form or another, and these participants were more sceptical about antibiotics. The accounts of those who experienced antimicrobial resistance were full of frustration and confusion because either the antibiotics had not worked or the first-line treatment had not worked. Likewise, in the qualitative arm of their study, Gaarslev et al. 92 established a growing number of patients who knew that antibiotics did not kill viruses and who agreed that taking antibiotics when not needed means that they may not work in the future. Although compliant with antibiotic treatment, participants in our study raised important questions of the right antibiotics being prescribed at the right time. Their accounts of illness suggested explicit and informed choices behind the experiences of both treatment and consultation for infections.

Participants tried to justify the prescription by relating to prescribers’ decision-making processes and their own and prescribers’ expectations. Some interviewees believed that prescribers seek to meet patient expectations who, in their turn, acted on an adherence principle. We suggest that adherence is often determined by complementarity of expectations. Although we did not attempt to quantify findings, we established that the rate of patients expecting antibiotics was relatively high (half of the participants). This finding, however, can be explained by a high proportion of patients who consulted for UTIs: an often recurrent health condition requiring appropriate antibiotic treatment. Our data also showed that patients could form expectations of expectations, trying to read the prescribers’ intentions and reflect on the dependency between what prescribers and patients wanted. Patients were also aware of the possibility of patient pressure. A patient acknowledged that other patients’ intentions may be based on exaggeration to receive a prescription. These accounts evidenced the observation that practitioners’ perceptions of patient expectations matter, rather than patient expectations per se.

Where participants were debating the appropriateness of prescribing, it was associated with informed choice and shared decision-making. Information sharing is a prerequisite to shared decision-making and it appears that patients want information about their medical condition and treatment options without necessarily being responsible for making treatment decisions. 106,107 In the medical sociological literature on late modernity, it is argued that individuals experience self, the body and the social and physical worlds with a high degree of reflection, questioning, evaluation and uncertainty. 108 It is assumed that the ‘consumerist’ patient and the ‘reflexive’ actor ‘both are understood as actively calculating, assessing and, if necessary, countering expert knowledge and autonomy with the objective of maximizing the value of services such as health care’. 108 Recently, it has been shown that when informed about individual and social consequences of antibiotic overuse, patients may be more receptive to antibiotic prescription limits. 109 This evidence suggested that the patient role involves staying informed about the issue of antibiotic use and considering potential benefits and harms when making decisions about antibiotic use. We found that informed patients (of antibiotics and the associated risks) displayed more satisfaction with the consultation. An interview study in Australia110 also established that most consumers would accept the GP’s decision not to prescribe an antibiotic if it was clearly explained. Therefore, the re-emergence of the informed patient is inevitable in the era of antimicrobial resistance, in which expert knowledge of antibiotics is broadcast through the media and public health campaigns.

The recent research demonstrated that patients were unwilling to follow the prescriber’s recommendations blindly and wanted to know about the appropriateness of prescribing,96 and our study of expectations and experiences lent support to this. Patients seemed to have been more prepared to openly deliberate on prescribing decisions and their expectations were more explicit than they were previously, despite that trust in the clinician still had a major role to play. 96,101,111 Those participants who emerged as informed patients rejected a blind compliance. Indeed, patient expectations were because of disclosure [e.g. it was manifested in the consultations that used elicitation (of expectations) technique]. Expectation elicitation by clinicians, directly or indirectly (i.e. by running commentary), and their open communication appeared important for an ongoing clinician–patient relationship. 95,101,112,113 Instead of trying to read the patient’s mind, prescribers were, and must be, making the expectations apparent by asking about them. 114

Strengths and limitations

We established the variety of patients’ expectations, which, in some cases, attested to an unquestionable compliance and, in other cases, to reflexive accounts of expectations. A wish for the right antibiotic with no resistance and no side effects prescribed at the right time (for bacterial infection) confirmed the patients’ expectations of appropriate treatment. This paramount expectation, according to the patients interviewed, was actualised in consultations with prescribers. Moreover, the patient experiences appeared more nuanced, and patients had more knowledge of antimicrobial resistance and side effects of antibiotics than might be assumed. Prescribers might be reassured that their patients may be knowledgeable and accepting of the limitations of antibiotics.

Although we investigated the accounts of patients, we were unable to interview the health-care professionals who treated our patient participants; however, we believe that the credibility of the research was enhanced by conducting interviews with a group of prescribers (reported in the next chapter). Similarly, our aim was not to observe actual encounters, but rather to examine the expectational structures of patients. Some qualitative research endeavoured to compare the perspectives of professionals and patients (e.g. Courtenay et al. 88), and questioned the pressure originating from patient expectations. Our analysis demonstrated that the participants believed that patient expectations did not influence prescription [in half of the encounters because the antibiotic treatment was evident (UTIs) and in the other half because expectations were undifferentiated]. However, because we did not have the prescribers’ accounts, we could not draw conclusions about the actual impact of patient expectations. Ideally, expectations on both sides should be studied to answer the validity of the notion of expectations of expectations. The study was conducted in the context of general practices in England. Although attempts were made to recruit patients based on purposive sampling, the sample of our study was skewed towards white British female patients who consulted for UTIs. The perspective may differ with gender, age, ethnicity or socioeconomic differences. This affects the results in the form of patients explicitly expecting antibiotics and their appropriate prescribing. Follow-up studies should attempt to diversify the sampling by purposefully including patients of different backgrounds who consult for other bacterial infections. For example, interpretation of our data suggested that education level might be influential. Future studies should also aim to include patients with infections who were not prescribed antibiotics to evaluate their experiences of illness and consultation.

Theme 1: risk assessment

Theme 2: balancing treatment risks

Theme 3: negotiating decisions and risks

Patient and public involvement

The process of developing subthemes and themes was discussed at a PPI meeting. The preliminary findings were presented and members were invited to discuss emerging themes and to review selected quotes from the interview transcripts for relevance. Feedback included comments on patient expectations, patient pressure for antibiotics, and trust and communication with GPs leading to additional interpretation.

Main findings in comparison with previous research

The study describes primary care prescribers’ perceptions of safety and associated trade-offs in the context of reduced antibiotic prescribing. We identify three key themes with relevance to safety: (1) risk assessment, (2) balancing treatment risks and (3) negotiating decisions and risks. These accounts from primary care demonstrated variations in prescribers’ approaches to decision-making behaviour, including perceptions of risks associated with prescribing or not prescribing antibiotics and in the communication of these decisions and risks to patients.

Decision-making for appropriate antibiotic prescribing was informed by safety considerations. Guideline-concordant risk assessment was generally preferred to tacit clinical judgement based on informal heuristics, in line with previous research. 138 Confidence in prescribing can be contrasted with views that accentuate diagnostic uncertainty. 118,126 In complex or uncertain cases, resolution was usually in favour of antibiotic prescribing, but this was in the context of a secular shift to generally more restrictive antibiotic prescribing behaviour. The reduction imperative co-exists with liberal prescribing, which was influenced by low tolerance of risks and patient pressures. This corresponds with extant literature that identifies the co-existence of different prescribing behaviours, including antibiotic compromising, antibiotic delaying and antibiotic withholding. 132

Safety trade-offs emerged from the respondents’ perceptions of risk by lending support to recent qualitative research that reported the complexity of balancing risks of antibiotic prescribing in hospitals. 139 In addition to the anticipated benefits, respondents identified multiple risks associated with either prescribing or not prescribing antibiotics, so that the immediate and long-term adverse effects of prescribing, including antimicrobial resistance, were weighed against potential complications of non-prescribing, such as sepsis. These untoward consequences rendered risk a double-edge sword. In the theory of social systems, such a conundrum can be described by the distinction risk/danger rather than risk/safety, as there is no absolute safety in prescribing decisions and hence the other side of risk remains danger, not safety. 140,141 From Luhmann’s140 perspective, some distinctions are two-sided forms of ‘second-order’ observations where one side is actualised at any given moment, but both sides can be seen as equally relevant in the situation. Risk/danger represents such a form, which exemplifies the contingency associated with seemingly binary choices, but which in itself is actuality versus potentiality. Safety experts, according to this perspective, are ‘first-order’ observers who may not account for the mutuality of contingency (i.e. the other side is always present in the background). Boiko et al. 141 applied this understanding to the analysis of clinical risks associated with anticoagulant prophylaxis, where risks of thrombosis were complemented by dangers of contraindications (e.g. bleeding). In our situation of antibiotic prescribing, the ‘risk’ element is associated with antibiotic prescribing potentially resulting in antimicrobial resistance and drug side effects, whereas the other side (danger) can be actualised if non-prescribing is chosen and can become the actual risk through the complications, such as sepsis. We found variation in how the prescribers perceived this duality, with the safety argument contributing in both directions (i.e. prescribing and non-prescribing). In other words, professionals acting on ‘doing something’ were juxtaposed against ‘doing no harm’ concerns. The participants were able to distinguish between short- (e.g. side effects) and long-term (e.g. antimicrobial resistance and effect on doctor–patient relationship) trade-offs of prescribing. Antimicrobial resistance was generally viewed as a stand-alone long-term adversity now being encountered in daily practice. It is gaining in prominence in contrast to findings from the earlier qualitative studies,142,143 and now has a more personalised relevance and clinical significance than some recent reviews have suggested. 144

Respondents negotiated safety in dealing with patients by rendering medical decision-making more explicitly during consultations. Patient expectations were found to be changing, as were the strategies employed to manage them. There was an emerging consensus on strategies to reduce antibiotic prescribing, including patient education, improved self-management advice and delayed prescribing, supported by patient-centred communication emphasised in the other literature too. 145 At the same time, our study showed that communication was primarily centred on warning signs and on maintaining a clinician–patient relationship, rather than on the discussion of risks and benefits with patients. This is consistent with previous findings that explicit analysis of trade-offs is most often undertaken by physicians alone, rather than as part of a dialogue with patients. 146 More explicit risk communication might become a focus of the consultations for (bacterial) infections. Systematic review evidence suggests that shared decision-making reduces prescribing,147 and our study also found that both delayed prescribing148–151 and safety-netting appeared to be effective strategies of shared decision-making.

Strengths and limitations

The study provided a coherent analysis of the views of primary care prescribers. It drew on participants working in rural and urban settings and included a sample that was diverse with respect to professional training and years of experience. The size of the sample may not have been sufficient to distinguish differences in approach between groups with different professional training, but this could be explored further in future studies. However, the study may have reduced transferability to other settings beyond UK primary care or beyond high-income countries. The study is based on interviews with prescribers and may be prone to the limitations associated with qualitative studies. Participants were necessarily informed of the nature and purpose of the research; consequently, both their participation in the interview and the interview responses might have been influenced by research participation. It is possible that respondents who were less inclined to reduce antibiotic prescribing might have been less prepared to participate. Interviewees might have been inclined to give what they perceived as ‘socially acceptable’ responses. We employed a thematic analysis because this enables a flexible investigation of a complex topic without drawing on pre-existing theory. To reduce the possibility of inconsistency, we employed a systematic, staged approach to analysis and a sample of transcripts was repeat coded by a second analyst.

Implications for further research

This study explored and characterised primary care prescribers’ perceptions of safety issues and risk management strategies relevant to reduced antibiotic prescribing. The study is a valuable investigation of primary care prescribers’ perceptions and, as such, it emphasises the safety perspective within the current debate on antibiotic prescribing and antimicrobial stewardship. The study identified dilemmas that are recognisable in the course of daily primary care practice and can form the basis for future improvement and antimicrobial stewardship programmes. Our research paves the way for a cross-sectional survey of risk perceptions. It highlights the need for further development of risk stratification and risk communication tools, such as decision-making checklists and evidence-based support tools. It also stresses the need for adequate training on antimicrobial resistance and reducing antibiotic prescribing (e.g. GRACE-INTRO and REDUCE). 152,153 Safety-netting had a strong presence in the interviews; however, this is under-researched and requires further exploration. Our findings support the argument136 that prescribers need more time to discuss the benefit–harm trade-off in shared decision-making, as this may help to reduce antibiotic prescribing in primary care.

Overall trends in antibiotic prescribing and infection recording

Figure 5 shows changes over time in age-standardised antibiotic prescribing rates per 1000 patient-years for coded and not coded indications. During the initial period of the study, from 2002 to 2012, the age-standardised total antibiotic prescribing rate per 1000 patient-years increased from 2002 (male, n = 423; female, n = 621) to 2012 (male, n = 530; female, n = 842), before declining in 2017 (male, n = 449; female, n = 753). The recent decrease in total antibiotic prescribing was accompanied by a decline in antibiotic prescribing for coded indications, but antibiotic prescriptions that were not associated with specific coded indications continued to increase. There was evidence of a decline in antibiotic prescribing for respiratory illness from 2008 onwards (Figure 6), and after 2012 there was evidence of decreasing prescribing for genitourinary and skin infections, as well as other specific indications. From 2002 to 2017, antibiotic prescriptions associated with non-specific codes increased, as did repeat prescriptions. Antibiotic prescriptions that were not associated with medical codes declined initially, but then remained constant (see Figure 6).

FIGURE 5.

Age-standardised antibiotic prescribing rates per 1000 patient-years for males and females from 2002 to 2017. (a) Male; and (b) female.

FIGURE 6.

Age- and sex-standardised antibiotic prescribing rates per 1000 patient-years for coded and not coded indications from 2002 to 2017. (a) Coded; and (b) not coded.

Variation in antibiotic prescription and infection recording among general practices

Table 8 summarises variation in antibiotic prescribing metrics between general practices in the sample. The 95% range for general practice-specific antibiotic prescribing rates was from 430 to 1038 antibiotic prescriptions per 1000 person-years, with a median of 648 antibiotic prescriptions per 1000 patient-years. The 95% range for the proportion of repeat prescriptions was from 3% to 24%. The 95% range for the proportion of antibiotic prescriptions with specific coded indications recorded ranged from 10% to 75%. Figure 7 presents a correlation matrix of the estimated general practice-specific random intercepts for different types of antibiotic prescribing. General practices with more antibiotic prescribing for RTI also tended to be higher prescribers for skin infections, genitourinary infections and other specific infections. Antibiotic prescribing for these conditions was, as expected, negatively correlated with prescribing for non-specific and non-coded indications, which were in turn correlated with each other.

FIGURE 7.

Correlation matrix of general practice-specific random effects for different measures of antibiotic prescribing. Red indicates positive correlation and blue indicates negative correlation.

Comparison of antibiotic prescribing and prescription recording in Vision and EMIS general practices

We evaluated whether or not similar results would be obtained in EMIS data from the CPRD Aurum database. In the CPRD Aurum sample, there were 158,305 participants from 883 general practices with 101,360 antibiotic prescriptions during 2017. In the CPRD GOLD sample, there were 160,394 patients from 290 general practices with 112,931 antibiotic prescriptions during 2017. This included 112 general practices in England and 178 in Scotland, Wales and Northern Ireland. The age- and sex-standardised antibiotic prescribing rate was 512.6 (95% CI 510.4 to 514.9) per 1000 person-years in CPRD Aurum and 584.3 (95% CI 582.1 to 586.5) per 1000 person-years in CPRD GOLD. The rate for CPRD GOLD practices in England was 505.2 (95% CI 501.6 to 508.9) per 1000 person-years, which is similar to the rate observed in CPRD Aurum.

Figure 8 presents age- and sex-specific antibiotic prescribing rates for 2017. Antibiotic prescribing was higher in children aged < 5 years, it decreased until the teenage years, and increased again (especially in women), before increasing steadily into older ages. This pattern of association was observed in both CPRD Aurum and CPRD GOLD, but estimates for CPRD GOLD were slightly higher than those for CPRD Aurum, but broadly similar when restricted to CPRD GOLD general practices in England. The lower panels of Figure 8 (see Figures 8b and 8d) provide Bland–Altman plots that presents the difference (95% CI) between all CPRD GOLD and CPRD Aurum practices (blue) and CPRD GOLD practices in England only (orange). For men and women in all age groups, CPRD GOLD general practices generally had slightly higher antibiotic prescribing rates than CPRD Aurum, whereas CPRD GOLD general practices in England had broadly similar antibiotic prescribing rates to CPRD Aurum. CIs were compatible, with no difference in any except the oldest age group (i.e. ≥ 85 years), for whom data are more sparse.

FIGURE 8.

Antibiotic prescribing rates in CPRD Aurum and CPRD GOLD, by age group and sex. Upper figure shows antibiotic prescribing rate per 1000 patient-years and lower figure shows difference in antibiotic prescribing rate per 1000 patient-years. (a) Antibiotic prescribing rate by age in CPRD Aurum and CPRD GOLD (male); (b) difference between all CPRD GOLD and CPRD Aurum practices (male); (c) antibiotic prescribing rate by age in CPRD Aurum and CPRD GOLD (female); and (d) difference between all CPRD GOLD and CPRD Aurum practices (female).

Patient and public involvement

The main findings from these analyses were discussed at a PPI meeting. Discussion focused on the reason for the increase in antibiotic prescribing between 2002 and 2012, whether or not further reductions are achievable, whether there are international differences or consensus in terms of antibiotic prescribing, and how the outcomes of antibiotic prescribing compared with no antibiotic prescribing could be compared.

Principal findings

This study found that antibiotic prescribing increased from 2002 to 2012, but declined subsequently, with changes over time being of larger magnitude for women than for men. The decline in antibiotic prescribing was earliest and most pronounced for RTIs, followed by other specific coded indications. We did not find evidence for a decline in antibiotic prescriptions, with poorly documented reasons for prescription.

We observed that nearly half of antibiotic prescriptions were not associated with specific coded indications, which is consistent with previous studies. 8,10 This suggests that total antibiotic prescribing is the most appropriate exposure measure for consideration in the following studies, because indication-specific antibiotic prescribing may be associated with considerable misclassification. Measures are needed to improve the recording of infection episodes in primary care both when antibiotics are prescribed and when antibiotics are not prescribed. Repeat prescriptions account for a significant proportion of uncoded prescriptions and repeat prescriptions might be indicated for prolonged or serious infections. Certain conditions may be associated with a higher rate of repeat antibiotic prescribing if there is initial treatment failure. For example, surgical intervention may eventually be required for treatment of empyema, osteomyelitis or infective endocarditis.

Comparison of Vision and EMIS data

Electronic health record systems may offer users discretion over the recording of data items. We found that EMIS data included similar proportions of antibiotic prescriptions being associated with no codes, non-specific codes and codes for infection episodes. This analysis suggests that antibiotic prescribing estimates from EMIS-derived data in CPRD Aurum are broadly similar to those obtained through analysis of Vision-derived data in CPRD GOLD. This similarity includes the rates of antibiotic prescriptions for subgroups of age and gender, the drug name and strength of antibiotic products prescribed, and the recording of medical diagnoses on the same day as the antibiotic prescription. We noted that antibiotics were more frequently prescribed in CPRD GOLD than in CPRD Aurum, but this was not the case when the CPRD GOLD sample was restricted to general practices in England. This suggests that antibiotic prescribing may be higher in Wales, Northern Ireland and Scotland, where prescription charges have been abolished since 2007, 2010 and 2011, respectively,156,157 although our study did not investigate the reasons for this difference. As well as slight differences in overall rates, we noted that drug choice might vary between databases. Trimethoprim prescribing was higher in CPRD GOLD than in CPRD Aurum. Nitrofurantoin has been recommended by Public Health England as the drug of first choice for UTIs in adults2 because of increasing antimicrobial resistance to trimethoprim, but this guidance may not apply in the devolved administrations. CPRD GOLD general practices in England were more similar to CPRD Aurum general practices with respect to prescribing of trimethoprim and nitrofurantoin. It is likely that differences in clinical practice between England and the devolved administrations (i.e. Scotland, Wales and Northern Ireland) may be greater than the differences between EMIS and Vision practices within England.

A strength of our study is that we used real-world data from primary care to estimate the rates of antibiotic prescribing. Using data from primary care is likely to provide a reliable picture of prescribing patterns, given that about 80% of all antibiotic prescribing in the UK’s NHS take place in primary care. 2 We estimated the difference between CPRD Aurum and CPRD GOLD, as well as the difference between CPRD Aurum and CPRD GOLD in England. A comparison between CPRD GOLD practices in England and CPRD Aurum practices was essential to benchmark recording in the CPRD Aurum database, which at the time of this study comprised contributing practices from England only. We recommend that future comparative studies of antibiotic prescribing in the UK should separately evaluate prescribing in CPRD Aurum compared with CPRD GOLD practices in England, and CPRD GOLD practices in Wales, Scotland and Northern Ireland, given that factors such as socioeconomic differences, as well as abolished prescription charges, in the devolved nations may have an impact on prescribing in these countries. There were generally only small differences between CPRD Aurum and CPRD GOLD in England. We acknowledge that we could have obtained greater precision with larger samples, but the present approach was pragmatic and provided sufficiently precise estimates for age-specific rates. We did not employ null-hypothesis significance testing, but elected to present CIs so that readers could reflect on the substantive importance of any estimated differences for their proposed studies. The community of CPRD researchers collectively has wide experience of compiling code lists for research in the CPRD GOLD database. Less experience is available for the CPRD Aurum database. We noted that CPRD Aurum product codes may be up to 17 characters in length, and the use of special programming features, such as the ‘bit64’ package in R, is required to maintain data integrity. We completed extensive searches for product codes to identify antibiotic products. We identified a larger number of potential products from the CPRD GOLD data dictionary, but a generally similar number of antibiotic product codes were actually recorded in the two data sets during 2017. Searches in the CPRD Aurum product dictionary should be based on term, drug substance and product names, as the BNF classification is less widely available in the CPRD Aurum product dictionary than in the CPRD GOLD. It may also be possible to compare ‘dm+d’ codes from the dictionary of medicines and devices, which are now employed in both CPRD GOLD and CPRD Aurum product dictionaries. We mapped medical code sets between CPRD GOLD and CPRD Aurum by matching on Read codes to make a like-for-like comparison. The analysis shows that, for these conditions, the use of the same Read codes gives similar results in CPRD Aurum and CPRD GOLD. There are some medical codes that are employed in EMIS only, which might be omitted through this process, and this merits further evaluation. Experience shows that, in Read-coded data, the majority of events are associated with a small number of codes; consequently, the omission of infrequently used codes is seldom important. Our main findings with respect to medical codes were consistent between databases. We also note that records of antibiotic prescribing do not indicate whether or not medicines were dispensed, whether or not they were taken, or whether they were taken by the patient they were prescribed to or by someone else. It is also possible that prescriptions recorded at out-of-hours visits, home visits or during attendance at residential care homes may be missing from the patient electronic record. Finally, the analysis undertaken was cross-sectional in nature and does not provide evidence about trends in antibiotic recording over time between CPRD GOLD and CPRD Aurum. We used data from the July and October 2019 releases of CPRD Aurum and CPRD GOLD, respectively; however, it may be preferable to compare the same month’s releases, and data for 2017 should be complete by 2019.

This study found that the analysis of EMIS-derived data in CPRD Aurum gives broadly similar estimates for antibiotic prescribing and infection recording to those reported for Vision-derived data in CPRD GOLD. CPRD GOLD includes general practices in Scotland, Wales and Northern Ireland, which have slightly higher antibiotic prescribing than either EMIS or Vision general practices in England. Based on these results, we believe that future research studies can be conducted in CPRD Aurum, informed by previous results from CPRD GOLD or The Health Improvement Network. It may also be possible to combine data from CPRD GOLD English practices with CPRD Aurum data for research on antibiotic prescribing. As CPRD Aurum includes an increasing number of general practices, this database will become increasingly important for public health research. However, further work is required to better understand the quality and completeness of information recorded in areas such as dosing regimen and treatment duration, which are important in estimating treatment exposure in pharmacoepidemiology and pharmacovigilance research.

Principal findings

The incidence of serious bacterial infections in men and women rose steadily between 2002 and 2017, particularly for sepsis (men and women), osteomyelitis (mainly in men) and kidney infections (mainly in women). The research aimed to test the hypothesis that patients from general practices with lower utilisation of antibiotics might have a greater risk of serious bacterial infections. We evaluated the incidence of serious bacterial infections, including 10 groups of infections that affect different systems of the body, as well as sepsis (including septicaemia). We did not find evidence that general practices that prescribe antibiotics less frequently might have a higher incidence of serious bacterial infections. We found evidence that each type of serious bacterial infection was recorded more frequently at general practices that record diagnostic codes for a high proportion of antibiotic prescriptions, suggesting that variation in the incidence of serious bacterial infection among general practices may be partly an artefact of data recording.

We conducted analyses after excluding repeat prescriptions, and these analyses raised the possibility that general practices with lower acute (non-repeat) antibiotic prescribing might have higher incidence of lung abscess and empyema and septic arthritis. However, these analyses were not pre-planned and should be considered as hypothesis generating, requiring confirmation in future studies. The incidence of these two conditions is less than 1 per 10,000 patients per year, and a relative rate of 0.9 for a 20% increase in prescribing implies that, at most, one additional case might arise every 10 years from a 20% reduction in prescribing at a general practice with 10,000 registered patients.

Strengths and weaknesses of the study

The study drew on data for a large population, comprising data for about 7% of the UK general population. In view of sample size constraints, antibiotic utilisation was estimated through analysis of data for a sample of patients, using hierarchical (multilevel) regression models to obtain general practice-specific antibiotic prescribing estimates. This contrasts with our previous study28 in which age- and sex-standardised rates were calculated from the data for each practice. Use of a regression modelling approach enabled us to make optimal use of the data, as well as adjusting for covariates that are associated with variations in antibiotic prescribing,159 including comorbidity, deprivation, region and calendar year, in addition to age and sex. 160 Consistent with previous studies,8,10 we observed that nearly half of antibiotic prescriptions were not associated with specific coded indications. This suggests that total antibiotic prescribing is the most appropriate exposure measure for consideration, as indication-specific antibiotic prescribing may be associated with considerable misclassification. Serious bacterial infections were identified from medical diagnostic codes recorded into primary care electronic health records, which include general practice records of consultations, hospital referrals and discharges. Many studies have shown that these records have a high predictive value for a range of diagnoses,161 but relying on a single data source can lead to underestimation of the total number of events. 162 CPRD records are linked to HES, but for only a subset of general practices in England, leading to a reduced sample size. There may be changes over time in the use of diagnostic categories, which might, in part, account for increasing diagnoses of ‘sepsis’. A study of US hospitals’ data found that there was a 706% increase in sepsis between 2003 and 2012, without any corresponding increase in positive blood cultures. 54 There was also an apparent increase in resistant infections, but this might also be due in part to data-recording changes and growing awareness of the problem of antimicrobial resistance, as well as true increases in resistant infections. An interrupted time series analysis158 offers an alternative approach to analysis, but this might be susceptible to changes over time in unmeasured confounders, such as code selection. The results of our study draw attention to the problem of poor coding quality in the context of infection management in primary care. Evidence from other studies suggests that missing values are typically missing not at random and the act of data recording may introduce a form of confounding by indication that may bias results. 163 To allow for this, we explicitly evaluated the extent to which differences in data recording between practices might account for variations in the incidence of serious bacterial infections. It is likely that misclassification of exposure and outcome variables from incomplete data recording might lead to underestimation of associations, although the direction of bias cannot always be anticipated. 164 We adjusted for a summary measure of comorbidity. Our analyses do not exclude the possibility that there may be vulnerable subgroups of patients, such as those with immunosuppression, who may be at increased risk if antibiotics are withheld.

Comparison with other studies

Consistent with our findings, Balinskaite et al. 158 reported increasing rates of infection in English primary care and hospital admissions data from 2010 to 2017. Their time series analysis suggested that antimicrobial stewardship intervention in 2015 had no impact on bacterial infections overall, but there was some evidence for increasing hospital admissions for peritonsillar abscess, decreasing hospital admissions for pyelonephritis and decreasing GP consultation rates for empyema. In a previous study, we found that PTA and pneumonia might be more frequent when general practices prescribe antibiotics less frequently for RTIs. 28 We did not include pneumonia in this study because we found that syndromes of ‘chest infection’ and ‘pneumonia’ may be difficult to distinguish in primary care records, with evidence of code shifting between the two categories. 165 In the present study, the incidence of PTA was not associated with total antibiotic prescribing. Randomised trials suggest that antibiotics protect against PTA166 and so it is plausible that this condition might be associated with respiratory antibiotic prescribing, but not with total antibiotic prescribing.

Incidence of peritonsillar abscess

There were 11,007 patients from 718 general practices in the May 2018 release of CPRD GOLD, with first episodes of PTA from 2002 to 2017. Fifty-three per cent of patients were male and 75% were aged 15–44 years (Table 12). The age-standardised incidence of new episodes of PTA was 17.2 per 100,000 patient-years for men and 16.1 per 100,000 patient-years for women. Between 2002 and 2017, annual age-standardised incidence rates ranged between 12.7 and 19.5 per 100,000 patient-years for both men and women, with no consistent trend. Age-specific incidence rates revealed that the incidence of PTA increased up to 15–19 years (men: 45.3 per 100,000 patient-years, 95% CI 44.7 to 45.8; women: 51.8 per 100,000 patient-years, 95% CI 51.3 to 52.4) and then decreased with increasing age (Figure 12).

FIGURE 12.

Age- and gender-specific 30-day probability of PTA after RTI consultation in patients who consulted for self-limiting RTI, with and without antibiotic prescription. Vertical lines represent 95% UIs. (a) Female; and (b) male.

Patients with PTA were more likely to be smokers (34% of PTA patients were current smokers vs. 16% of person-time in a sample of the general population), whereas 12% of both groups had previously smoked. This effect was consistent when stratifying by gender and remained consistent over time. The prevalence of comorbidity was lower among PTA cases (28.6%) than in the general population (40.8% of all follow-up time was in patients with comorbidities). This pattern persisted when stratifying by gender. When stratifying by age, comorbidity prevalence was lower in PTA patients than in the general population in people aged 15–64 years, but not in people aged < 15 years or ≥ 65 years. PTA was evenly distributed across general practice IMD deciles (see Table 12). There were slightly more cases in winter and spring than in summer and autumn.

Consultation and prescription patterns before peritonsillar abscess diagnosis

Among 11,007 incident cases of PTA, 6996 patients (64%) had consulted the general practice in the 30 days preceding diagnosis of PTA, including 4243 (39%) consulting with any self-limiting RTI (Table 13). The majority of consultations were in the 7 days before PTA diagnosis. Similar patterns of associations were observed when considering 14- and 60-day thresholds. There were 3782 (34%) patients who received a prescription for antibiotics in the 30 days before PTA diagnosis.

Decision tree

A decision tree was constructed to calculate the 30-day probability of RTI consultation, antibiotic prescription and PTA. For the whole population, the probability of consultation for self-limiting RTI in a 30-day period was 0.0275 (95% CI 0.0275 to 0.0276). The probability of antibiotic prescription at a self-limiting RTI consultation was 0.483 (95% CI 0.483 to 0.484). Age- and gender-specific probabilities of PTA, given that patients had consulted for RTI, were also calculated (Table 14). The probability of PTA was higher in men than in women and was the greatest between 15 and 34 years of age (see Figure 12). In both men and women and at all ages, the probability of PTA following a RTI was lower if an antibiotic was prescribed. The NNT with an antibiotic to prevent one case of PTA was also calculated for each age group and gender. The NNT was lower for patients with the highest probability of PTA. At aged 25–34 years, the number of patients with RTI who would need to be prescribed an antibiotic to prevent one PTA case was 926 (95% CI 814 to 1063) for men and 1984 (95% CI 1756 to 2263) for women. The numbers were similar in men and women aged 15–24 and 35–44 years, but were substantially greater for either younger or older patients (Table 15).

Linked hospital episodes data

In the subset of patients with linked hospital records, the age-standardised incidence of PTA was 20.6 per 100,000 patient-years for men and 19.0 per 100,000 patient-years for women. The distribution of cases over time and between age groups was similar between CPRD and HES data (see Report Supplementary Material 3, Supplementary Figure 1). Estimates of the risk of PTA with and without antibiotics were consistent with data from CPRD. The protective effect of antibiotics was also evident when considering only PTA cases resulting in hospitalisation (see Report Supplementary Material 3, Supplementary Figure 2).

Patient and public involvement

The findings from these analyses were presented at a PPI meeting. Patients observed that it would be useful to understand the experience of having PTA and this would help to clarify the nature of any benefit from antibiotic prescribing. However, benefits from prescribing were at the individual level, whereas from the population perspective the risks from antimicrobial resistance might be more concerning. In the worst-case scenario, reduced use of antibiotics would have to be tolerated to ensure that antibiotics did not stop working. Patients commented that the results were an ‘eye opener’, with GPs needing to see these numbers before thinking about prescribing. The results could also be communicated to patients to inform their understanding of when antibiotics might be needed.

Principal findings

There have been few large population-based studies of PTA. 167 This study provides precise estimates for disease incidence and showed that PTA is most frequent in young people aged 15–24 years and is more frequent in men. Nearly two-thirds of patients consulted at their general practice during the 30 days before PTA diagnosis, with about half of these consultations being recorded as self-limiting RTIs. The analyses demonstrate a protective effect of antibiotic prescribing against PTA following a RTI consultation for both men and women and in all age groups. However, the overall risk of PTA was low, and ≥ 1000 antibiotic prescriptions would be required to prevent one PTA, even in the age group at the highest risk. Consultation rates for respiratory illness are generally considerably higher in women than in men, and this may account for the generally lower risk of PTA following a consultation for RTI in women than men.

Strengths and weaknesses of the study

The study benefited from a very large sample size drawn from all parts of the UK. The study population comprised just under 10 million patients, with 11,000 PTA cases over 16 years, providing precise estimates for a rare outcome. A key limitation of this study is that its non-randomised design meant that patients were not randomly assigned to antibiotic treatment. It is likely that patients who received antibiotics had more severe illness and a greater risk of PTA than patients who did not receive antibiotics. Consequently, the risk of PTA in antibiotic-treated patients may have been overestimated, and that in untreated patients underestimated, in comparison with possible random allocation. As a result, the protective effect of antibiotic treatment may have been underestimated and the true number of antibiotic prescriptions needed to prevent one PTA may be smaller than we estimate. It was not possible to include data on symptoms or severity of RTIs. We searched patients’ records for markers of illness severity, including fever, lymphadenitis or mention of Centor criteria,170 but these were recorded in only a very small number of cases. As noted by others, not all consultations in primary care electronic records have a clinical code assigned to them. 8 Almost one-third of PTA patients consulted their general practice with a self-limiting RTI in the 30 days preceding their PTA diagnosis, but this figure may be an underestimate because it is known that acute illness presentations are not always well coded in primary care. 8 The data for the study were recorded prospectively by family physicians and other primary care professionals, but the data were collected for clinical practice and not research. Standardised recording procedures were not used, leading to variability in data recording. The estimates represent values expected at an average practice, but estimates might vary depending on the consultation and prescribing rates or data recording procedures used at different practices. Previous studies have shown that diagnoses recorded into electronic health records have high predictive validity,161 but there is evidence of differential recording in alternative data sources. 162 Our analyses showed that analysis of hospital records, in addition to primary care records, did not change conclusions overall. In a clinical trial, an overall effect is usually estimated before subgroup analyses are performed, but this approach is not always desirable in epidemiological analysis of population data, especially when rates vary considerably by age. 171,172 In view of the age-related variations in incidence of PTA, we analysed data by subgroup of age and gender, but we note that small differences across subgroups should be interpreted with caution.

Comparison with previous studies

In a previous study, there was evidence that general practices with lower rates of antibiotic prescription for RTI had a higher incidence of PTA,28 it being estimated that there might be one additional case of PTA over 10 years for a general practice that reduced antibiotic prescribing for RTI by 10%. 28 The present individual patient-level analyses show that antibiotic prescription at a RTI consultation is associated with reduced risk of PTA, consistent with the more limited evidence from randomised clinical trials. In a review of randomised controlled trials,166 the relative risk of PTA within 60 days of a sore throat was 0.15 (95% CI 0.05 to 0.47) after antibiotic treatment compared with no antibiotic. However, most evidence was provided by one trial published in 1951 with a relatively small sample size. 166,173 Previous non-randomised studies also suggested a protective effect of antibiotics against PTA, but did not quantify age- and gender-specific risks following a primary care consultation for RTI. 56,58,174

Our estimates of PTA incidence (16–17/100,000 person-years) from CPRD data were consistent with those of previous studies,28 although our study spanned a wider time period. Previous studies have been limited by not including data from hospital discharge records, but we found that the associations between PTA incidence and antibiotic use were consistent whether we analysed CPRD data only or also included HES data. Consistent with previous studies, our data show an excess of current smokers compared with population reference data for smoking prevalence. 175 Comorbidity is not usually considered a risk factor for PTA. 167 Consistent with Klug’s review,167 there was only weak evidence of seasonality.

The risk of PTA might be higher following presentations for sore throat or tonsillitis. However, respiratory infections often have mixed presentation, beginning with a sore throat, followed by symptoms of coryza and later cough. 176 RTI presentations are also often indistinctly coded, making analysis of subgroups challenging. 8 Clinical guidelines for the management of sore throat in the UK advocate the use of clinical risk scoring with either Centor170 or FeverPAIN177 scores. Bacteriological testing is discouraged and is infrequent in the UK, possibly in contrast to the USA, where this is endorsed in some clinical guidelines. 178

Main results

The study included 706 general practices with a total of 66.2 million person-years of follow-up (Figure 13). Data for the distribution of sepsis patients by age and gender are shown in Table 16. The probability of a consultation with a common infection of the skin, respiratory tract or urinary tract in any 30-day period ranged between 0.02 (1 in 50) and 0.08 (1 in 12). This probability of an infection consultation was greater in children and older people and higher in women than men during mid-life (Table 17). The probability of an antibiotic being prescribed at an infection consultation ranged from 0.43 to 0.67, being lowest for young children, in whom consultation rates are highest (see Table 17).

FIGURE 13.

Flow chart showing participant selection for the main and linked samples.

There were 35,244 first episodes of sepsis between 2002 and 2017. The probability of an infection consultation within 30 days before a sepsis event ranged between 0.14 (one in seven) and 0.26 (one in four), with larger values at the extremes of age (see Table 17). If no antibiotic was prescribed, the probability of sepsis at age 0–4 years was 0.000054 (1 in 18,519 consultations) in males and 0.000060 (1 in 16,667 consultations) in females. The probability of sepsis following an infection consultation without antibiotics increased linearly with age on a log-scale (Figures 14a and b), reaching 0.004647 (1 in 215 consultations) in males and 0.003110 (1 in 321 consultations) in females aged ≥ 85 years (see Table 17). If antibiotics were prescribed at an infection consultation, the estimated probability of sepsis was lower, ranging from 0.000020 (1 in 50,000 consultations) in males to 0.000023 (1 in 43,478 consultations) in females aged 0–4 years to 0.000833 (1 in 1200 consultations) in males and 0.000509 (1 in 1965 consultations) in females aged ≥ 85 years. The number of antibiotic prescriptions required to prevent one sepsis event was highly age dependent (see Figures 14c and d). Among children aged 0–4 years, the NNT was 29,773 (95% UI 18,458 to 71,091) in males and 27,014 (95% UI 16,739 to 65,709) in females. However, at the age of ≥ 85 years, the NNT was 262 (95% UI 236 to 293) in males and 385 (95% UI 352 to 421) in females.

FIGURE 14.

Probability of sepsis following infection consultations in primary care if antibiotics are prescribed or not for (a) males and (b) females. Number of antibiotic prescriptions required to prevent one sepsis event (i.e. NNT) for (c) males and (d) females. Figures are median probability (95% UI).

In the population aged ≥ 55 years, estimates were obtained separately by frailty category (Figure 15). The probability of sepsis was greater and the NNT smaller for patients with more advanced frailty. For ‘non-frail’ patients aged 65–74 years, the NNT was 1680 (95% UI 1354 to 2133) for men and 2718 (95% UI 2089 to 3697) for women. However, for patients of the same age with severe frailty, the NNT was 259 (95% UI 196 to 360) for men and 438 (95% UI 329 to 624) for women. For patients with severe frailty, the NNT was < 250 in men and < 400 in women for all age groups over 55 years. For non-frail patients, the probability of sepsis increased and the NNT decreased with increasing age (see Figure 15). In ‘non-frail’ patients, the NNT declined from 2309 (95% UI 1890 to 2879) in men and 3782 (95% UI 3001 to 4907) in women at 55–64 years to 407 (95% UI 274 to 677) in men and 499 (95% UI 346 to 780) in women aged ≥ 85 years. Estimates for patients with ‘mild’ or ‘moderate’ frailty exhibited an intermediate pattern (see Figure 15).

FIGURE 15.

Number of antibiotic prescriptions required to prevent one sepsis event (i.e. NNT) following infection consultations in primary care, by frailty level, gender and age group. Figures are median estimate (95% UI). (a) Male; and (b) female.

The probability of sepsis was higher following consultations for UTI than for skin infections or RTI; a pattern of association that was observed across all age groups and in both men and women (Figure 16). For patients aged 65 years without antibiotic treatment, the probability of sepsis following a RTI consultation was 0.00090 (1 in 1111 consultations) in men and 0.00053 (1 in 1887 consultations) in women. Following a skin infection consultation, the probability of sepsis following a RTI consultation was 0.00224 (1 in 446 consultations) in men and 0.00150 (1 in 667 consultations) in women. Following a UTI consultation, the probability was 0.009227 (1 in 108) in men and 0.003787 (1 in 264) in women. At the same age, the corresponding NNTs for RTI were 1257 (95% UI 1112 to 1434) for men and 2278 (95% UI 1965 to 2686) for women. For skin infection, this was 502 (95% UI 398 to 646) for men and 784 (95% UI 602 to 1051) for women. Finally, for UTI consultations, this was 120 (95% UI 102 to 145) for men and 284 (95% UI 241 to 342) for women (see Figure 16).

FIGURE 16.

Number of antibiotic prescriptions required to prevent one sepsis event (i.e. NNT), by age group, gender and type of infection consultation. Figures are median estimate (95% UI). UIs were omitted for age 0–4 and 5–9 years if data were too sparse to give reliable estimates. (a) Male; and (b) female.

Sensitivity analyses

An analysis employing a 60-day time window to evaluate exposure gave generally similar results to those using a 30-day time window. In men aged ≥ 85 years, the NNT for all infections was 262 (95% UI 236 to 293) with a 30-day time window, but 313 (95% UI 276 to 359) with a 60-day time window. For women of the same age, the NNTs were 385 (95% UI 352 to 421) and 466 (95% UI 419 to 523), respectively. When the analysis results were compared for the 4-year periods from 2002 to 2005 with from 2014 to 2017, estimates for the probability of sepsis were slightly higher and for NNT were slightly lower for the more recent period (Figure 17), consistent with the higher reported incidence of sepsis in this period. In the oldest age group (i.e. aged ≥ 85 years), in 2014–17, the probability of sepsis without antibiotics was 0.007287 in men and 0.004775 in women. The probability of sepsis with antibiotics was 0.001290 in men and 0.000839 in women. The NNT was 167 (95% UI 141 to 202) for men and 254 (95% UI 216 to 302) for women.

FIGURE 17.

Estimates for number of antibiotic prescriptions needed to prevent one sepsis episode (i.e. NNT) for four periods: 2002–5, 2006–9, 2010–13 and 2014–17. (a) Male; and (b) female.

In the linked sample, there were 42,785 first sepsis events, considering the first sepsis event in each patient across all three data sources: 17,341 from primary care records, 17,363 from HES Admitted Patient Care primary diagnoses and 8081 from ONS mortality records during 36.2 million patient-years’ follow-up. Accordingly, the underlying probability of sepsis was greater when linked records were employed. However, sepsis events in HES and ONS mortality statistics were less frequently associated with preceding infection consultations in general practice (see Report Supplementary Material 3, Supplementary Figure 3). Consequently, the probability of sepsis following an infection consultation was only slightly higher if linked data were included in the analysis (see Report Supplementary Material 3, Supplementary Figure 4) and the estimated NNT was only slightly lower (see Report Supplementary Material 3, Supplementary Figure 5).

Patient and public involvement

The results from these analyses were discussed at a PPI meeting. The results were considered to be informative. However, a wider range of influences, such as the patient’s past medical history, the severity of their symptoms and their expected tolerance of a condition, might also be important. However, general practice record-keeping is often very poor in respect of not only infection consultations and antibiotic prescribing, but also in other areas of medical practice. This could result from limitations of general practice information systems, as well as limitations of staffing, with poor practice often being attributed to locums. This might mean that the required information was not available for analysis. Furthermore, there is often wide variation between general practices so that average findings might not always be transferable to specific contexts.

Main findings

This study analysed primary care electronic health record data for a large population followed for 16 years with 35,244 new sepsis events. We found that the probability of sepsis following consultation for common infection episodes in primary care is highly age dependent. Without antibiotic treatment, sepsis may follow < 1 in 10,000 infection consultations for those aged < 25 years and < 1 in 1000 consultations for those aged < 65 years. The probability of sepsis increases at older ages, and sepsis may follow approximately 1 in 200 (men) or 1 in 300 (women) consultations at the age of ≥ 85 years. At older ages, the probability of sepsis is also highly dependent on frailty level (i.e. 55 year olds with severe frailty have a similar probability of sepsis as non-frail 85 year olds). The probability of sepsis is related to infection type, being greatest following consultations for UTI and least for RTI, and consultations for skin infections being in an intermediate position. Risks were generally slightly higher for men, which might be accounted for by their generally lower consultation rates.

The incidence of recorded sepsis has been increasing over time, with more inclusive case definitions and increasing awareness of the condition. 39,54 When we estimated the main results for the period 2014–17, the probability of sepsis was higher and the NNT was lower than for the period 2002–17. Although we caution that the absolute values of estimates may vary depending on the temporal or geographical context, we expect that in relative terms estimates will continue to identify older age, frailty and UTI consultations as being associated with greatest risks of sepsis.

Sepsis is an uncommon but concerning outcome of common infection episodes in primary care. Appropriate antibiotic therapy may have immediate benefits that are not restricted to reduction in risk of sepsis, but antibiotic prescriptions are also often associated with immediate harms in the form of drug side effects. The potential risk of antimicrobial resistance has a significance that extends beyond the context of an individual consultation. Prescribing decisions must, therefore, be informed by the balance of all of the benefits and harms of either prescribing or not prescribing antibiotics. Quantification of the possible risks of sepsis contributes to informing these decisions.

Strengths and limitations

The study drew on a large population-based cohort, enabling us to analyse representative data and obtain precise estimates that may be widely applicable. However, electronic health records comprise clinical data with several limitations and potential for bias. We analysed the outcomes of clinical decisions on whether or not to prescribe antibiotics. In the absence of randomisation, it may be expected that antibiotics were prescribed to individuals at higher risk, whereas lower-risk patients may be less likely to be prescribed antibiotics. Consequently, the probability of sepsis may be underestimated (in comparison with a study employing random allocation) in the absence of antibiotics and overestimated for patients receiving antibiotics, with the NNT being overestimated. However, the analysis depended on general practice electronic health records of antibiotic prescriptions, which account for about 85% of community antibiotic prescribing,2 but we cannot exclude the possibility that patients might have obtained antibiotic prescriptions from alternative sources, including out-of-hours services. Measures of illness severity are rarely recorded for common infection consultations in primary care and so it was not possible to adjust for illness severity in analyses. It is also established that not all infection consultations in primary care are correctly coded, leading to underestimation of consultation rates. 10 We included data from 706 general practices over a 16-year period. Infection consultation and antibiotic prescribing rates were estimated from sample data. The estimates in this paper represent average values for this population of general practices and period of time. However, we conducted a sensitivity analysis with data from 2014 to 2017 only. We also acknowledge that, in addition to changes in overall antibiotic utilisation, there have been changes in the proportion of prescriptions for broad-spectrum antibiotics. Future studies might be designed to compare the probability of sepsis if broad- or narrow-spectrum antibiotics are prescribed. The sample design used to estimate infection consultation rates and antibiotic prescribing proportions gave each practice and each study year equal weight, but we could have weighted the sample by practice size.

We analysed data for infection consultations in primary care and compared outcomes if antibiotics were or were not prescribed. However, previous studies have shown that antibiotics may be prescribed at consultations with no definite diagnosis recorded. 8,10 We did not include these prescriptions because there was no valid comparator in terms of consultations without antibiotic prescriptions, but conclusions might have differed if missing diagnosis information had been complete. We caution that the precise values of these estimates may be expected to vary in different local contexts and according to the types of infection circulating in a community at a given time. We did not employ the approach of null hypothesis significance testing and do not report p-values. We evaluated association modification by age, gender, frailty level and consultation type. We employed the eFI, which is a well-described measure based on 36 deficits,65 although we also applied it in the age range 55–64 years, in which it is less well documented. We estimated stratified values for broad groups of patients defined in terms of age, gender and frailty. We did not estimate personalised risks for individual patients and the clinical circumstances in each specific consultation should be used to inform estimates of sepsis risk for individuals. We relied on clinical records of sepsis from general practice, but we cannot be sure that all sepsis events were community rather than hospital acquired. In the UK, patients register with a general practice for continuing care, but patients may utilise emergency and out-of-hours services for acute problems, such as sepsis, and these events might not be captured in primary care records. Providers may vary in their use of the term ‘sepsis’, which is an intermediate condition linking an infection and organ damage consequent on infection. The selection of clinical terms and medical codes is at the discretion of clinical staff, leading to lack of data standardisation. The conditions identified as ‘sepsis’ may represent a range of disease severity, and probability estimates might be proportionately lower if only severe sepsis was included. However, by using linked data, we showed that inclusion of hospital episodes and mortality records did not lead to any important changes in conclusions. Further research is needed to refine, update and improve the accuracy of these initial estimates.

Comparison with other studies

There has been a trend towards more frequent recording of sepsis in recent years, and this has not always been accompanied by evidence of increased bloodstream infections. In an interrupted time series analysis, Balinskaite et al. 174 found no evidence that antimicrobial stewardship interventions in the UK might be associated with increased rates of sepsis. In an ecological analysis,39 we did not find evidence that general practices with lower antibiotic prescribing might have greater risk of sepsis over the same period of time and in the same practices as were included in the present study. Gharbi et al. 61 found that in older adults presenting with UTI, there was an increased risk of sepsis if antibiotic prescriptions were not given or were delayed. The present results extend these findings by estimating risks across all ages, different levels of frailty and different types of infection consultations. The lack of consistency between estimates from ecological- and individual-level analyses is likely to be explained by the substantial proportion of patients with sepsis who present without preceding infection consultations in primary care, as well as the small proportion of higher-risk consultations that are not associated with antibiotic prescriptions. RTI consultations are extremely frequent, which may account for the lower probability of associated sepsis. Respiratory infections are often caused by viruses, and clinicians may tend to reserve the term ‘sepsis’ for bacterial infections. We evaluated uncomplicated lower UTIs, but estimates for the probability of sepsis might have been higher if kidney infections had been included.

Incidence of sepsis in each data source

There were 4,081,214 registered patients from 378 general practices who were eligible for linkage and contributed 36,209,676 patient-years of follow-up. There were 20,206 patients with a first episode of sepsis recorded in primary care electronic health records, 20,278 in HES and 13,972 in ONS. The characteristics of patients with sepsis from each data source are presented in Table 18. Each data source showed a slightly higher proportion of female patients than male patients (CPRD, 51%; HES, 52%; ONS, 55%). The frequency of sepsis increased with age, with a maximum over 75 years of age. The most deprived IMD quintile had the highest proportion of sepsis cases for each data source for men and women (25–26%) and the least deprived quintile consistently had the lowest (15–17%). The number of first sepsis events increased over time for primary care and hospital records, with more first episode sepsis cases in the period 2012–17 for both males and females. In CPRD, the period 2012–17 accounted for 44% of sepsis episodes for men and 42% for women. In HES, the latest period accounted for 50% of cases for both genders. This increasing trend was not apparent in mortality records. The period with the largest number of sepsis deaths in the ONS file was 2007–11.

Annual age- and gender-standardised incidence of sepsis is shown for each data source in Report Supplementary Material 3, Supplementary Figure 6. Primary care electronic records showed a steady increase in sepsis cases over the study period, with an acceleration from 2012 to 2017. In the CPRD, in 2002, the age-standardised incidence was 0.35 (95% CI 0.32 to 0.39) per 1000 person-years in male patients and 0.34 (95% CI 0.30 to 0.37) per 1000 person-years in female patients. By 2017, this had increased to 1.15 (95% CI 1.04 to 1.26) per 1000 person-years among male patients and 1.10 (95% CI 95% CI 1.00 to 1.19) per 1000 person-years among female patients. Consistent with primary care data, HES data showed a steep increase in sepsis cases over the study period, particularly from 2012 to 2017. In HES records, for 2002, the age-standardised incidence was 0.29 (95% CI 0.25 to 0.32) per 1000 person-years among male patients and 0.25 (95% CI 0.22 to 0.28) per 1000 person-years among female patients. By 2017, this had increased to 2.52 (95% CI 2.35 to 2.68) per 1000 person-years among male patients and 2.10 (95% CI 1.96 to 2.23) per 1000 person-years among female patients. In contrast to primary care and hospital records, there was no consistent trend in the recording of sepsis as a cause of death during the study period. In 2002, sepsis mortality was 0.34 (95% CI 0.31 to 0.38) per 1000 person-years among male patients and 0.28 (95% CI 0.26 to 0.31) per 1000 person-years among female patients. By 2017, this had risen to 0.42 (95% CI 0.35 to 0.49) per 1000 person-years among male patients and 0.30 (95% CI 0.25 to 0.35) per 1000 person-years among female patients, although the highest peak in mortality rate was in 2006 for female patients and 2007 for male patients.

Concurrent recording of sepsis events across data sources

There were 4117 (20%) first HES sepsis events and 2438 (17%) mortality records that were concurrent with incident primary care sepsis records, based on a 30-day time window (Figure 18). Among the 13,972 deceased patients with sepsis listed as any cause of death, 2438 (17%) had an incident sepsis event recorded in primary care electronic health records and 3397 (24%) had incident sepsis events recorded in HES in the same period. Only 614 patients had index sepsis events recorded across all three data sources within 30 days of the date of event or date of death. We evaluated whether or not extending the time window influenced conclusions. Including both first and subsequent sepsis records and a 90-day time window gave 4770 (24%) HES records and 2635 (19%) mortality records concurrent with an incident primary care sepsis record.

FIGURE 18.

Incident first sepsis events in CPRD, HES and ONS using a 30-day time window to evaluate concurrence.

Variables associated with concurrent records

Report Supplementary Material 3, Supplementary Figure 7, shows the results of a logistic regression analysis of variables associated with concurrent recording in more than one data source for patients with sepsis. Among patients with sepsis events recorded in primary care, concurrent HES records of sepsis within 30 days before or after first diagnosis were lowest in the age range 5–34 years, but higher at younger or older ages. There was no consistent association with deprivation overall. There were higher odds of concurrent recording in the most recent period of diagnosis. For index sepsis events recorded in HES, concurrent recording in primary care electronic health records was not consistently associated with age, deprivation or period. Among patients with sepsis events recorded in ONS, there were lower odds of a concurrent CPRD record of sepsis for the age range 35–54 years, being registered with a practice in the least deprived quintiles and the most recent period of diagnosis.

Incidence of sepsis from linked data sources

Table 19 shows age-specific sepsis incidence rate per 1000 patient-years from primary care electronic health records and for primary care electronic health records combined with HES and ONS, stratified. With primary care electronic health records alone, sepsis incidence rates increased from 0.19 in male patients and 0.20 in female patients per 1000 person-years for those aged < 5 years to 5.22 in male patients and 3.55 in female patients aged ≥ 85 years. Estimated incidence rates were substantially higher when either HES or mortality records were included. When all three data sources were combined, the incidence of sepsis was 0.32 per 1000 person-years in male patients and 0.36 per 1000 person-years in female patients aged < 5 years, increasing to 10.09 per 1000 person-years in male patients and 7.22 per 1000 person-years in female patients at age ≥ 85 years. Report Supplementary Material 3, Supplementary Figure 2, shows equivalent age-specific incidence when first sepsis records in each calendar year, rather than the first in the study period, were included.

Description of themes

Four themes emerged from the interviews, relating to the decision to use the web pages. Subthemes were identified within each theme, which are presented in Table 20.

A decision-making tool

General practitioners discussed the use of the web pages as a decision-making tool. Most GPs considered that the pages would be useful for decision-making in practice. Some GPs thought that the pages would be useful for training and some discussed their use for less experienced staff.

Use with patients

General practitioners discussed using the web pages with a patient and how they would show these to a patient. However, some GPs felt that they would not use the pages with a patient at present.

Design of tool

General practitioners discussed the way in which the design of the pages could help them during a consultation and also how these could be improved. Overall, GPs felt that the pages provided a useful selection of data. However, GPs felt that clearer descriptions were needed and that the data should be simplified further.

Accessing pages during consultation

General practitioners discussed the ways in which the pages could be accessed during a consultation. Some GPs felt that the pages should pop up automatically, whereas others said that they would prefer it if pages popped up on their screen only if they had selected them.

Development of tool

The web pages were refined throughout the interview process based on continuing feedback. Early interviews provided many constructive criticisms and suggestions for change. The themes informed areas for change and extensive analysis of transcripts informed the exact amendments to be made to the pages to increase acceptability and aid usage. Key changes and adaptations made to the prompts and the main themes that informed these can be seen in Table 21.

Implications for health care

• The research found that antibiotic prescribing in primary care is decreasing, but the decline is most evident for prescriptions with clearly defined indications recorded. Incompletely coded prescriptions have not decreased. Improving the recording of infection episodes is important for informing antimicrobial stewardship in primary care.

• Both antibiotic prescribing and the coding of prescriptions vary widely between general practices. The research did not find evidence that general practices with lower total antibiotic prescribing might have more frequent occurrence of serious bacterial infections. Serious bacterial infections were more frequently recorded at general practices with higher proportions of informatively coded infection consultations.

• The research provided stratified estimates of risk that identify groups of patients in whom, and types of consultations in which, antibiotic prescribing can be more safely reduced. We developed an interactive app that can be used to communicate these estimates to primary care prescribers. We found evidence that serious bacterial infection complications were generally less frequent if antibiotics were prescribed, but the possibility of benefit depended on the underlying frequency of the complication in the context of a particular patient’s characteristics.

• The safety trade-offs associated with either use or non-use of antibiotics present difficulties, especially when prescribing decisions are inconsistent with patients’ expectations. The research highlighted how patients’ expectations are now more complex than earlier research reported and exhibit tensions between adherence to antibiotics and consuming antibiotics in more reflexive, informed ways. Ensuring that present and future patients are better informed about both the potential benefits and the potential harms of antibiotic use will contribute to future antimicrobial stewardship.

Recommendations for research

• Measures are needed to improve the recording of infection episodes in primary care both when antibiotics are prescribed and when antibiotics are not prescribed. Interventions should be developed and tested to improve the quality of infection recording in primary care electronic health records and ensure consistency of terminology and coding across primary and secondary care.

• Estimates for antibiotic prescribing and infection recording were broadly similar in both CPRD GOLD and CPRD Aurum databases, suggesting that future research on antimicrobial stewardship may be conducted using primary care data in CPRD Aurum.

• The conditions identified as ‘sepsis’ may represent a range of disease severity and further research is needed to refine the predictive accuracy of models of sepsis following primary care infection consultations.

• The app developed for this research should undergo further iterative development to incorporate antibiotic prescribing and coding information, drawn from individual patient data rather than the aggregate data presently utilised in existing information feedback strategies. This can then be employed as an antimicrobial stewardship tool and tested in a randomised controlled trial.

• Previous research into antibiotic prescribing practices in primary care may need to be updated to include the need to understand more about prescribing behaviour by professional background (e.g. GP/nurse/pharmacist), risk perceptions and further research on the quality of prescribing information and safety-netting by clinicians.

Preprints

Boiko O, Burgess C, Fox R, Ashworth M, Gulliford MC. Risks of use and non-use of antibiotics in primary care. Qualitative study of prescribers’ views. BMJ Open 2020;10:e038851. https://doi.org/10.1136/bmjopen-2020-038851

Gulliford MC, Sun X, Anjuman T, Yelland E, Murray-Thomas T. Comparison of antibiotic prescribing records in two UK primary care electronic health record systems: cohort using the CPRD GOLD and CPRD Aurum databases. BMJ Open 2020;10:e038767. https://doi.org/10.1136/bmjopen-2020-038767

Rezel-Potts E, Gulliford MC, Safe AB Study Group. Sepsis recording in primary care electronic health records, linked hospital episodes and mortality records: population-based cohort study in England. PLOS ONE 2020;15:e0244764. https://doi.org/10.1371/journal.pone.0244764